wisent 0.7.701__py3-none-any.whl → 0.7.901__py3-none-any.whl

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (330) hide show
  1. wisent/__init__.py +1 -1
  2. wisent/core/activations/activation_cache.py +393 -0
  3. wisent/core/activations/activations.py +3 -3
  4. wisent/core/activations/activations_collector.py +9 -5
  5. wisent/core/activations/classifier_inference_strategy.py +12 -11
  6. wisent/core/activations/extraction_strategy.py +256 -84
  7. wisent/core/classifiers/classifiers/core/atoms.py +3 -2
  8. wisent/core/cli/__init__.py +2 -1
  9. wisent/core/cli/agent/apply_steering.py +5 -7
  10. wisent/core/cli/agent/train_classifier.py +19 -7
  11. wisent/core/cli/check_linearity.py +35 -3
  12. wisent/core/cli/cluster_benchmarks.py +4 -6
  13. wisent/core/cli/create_steering_vector.py +6 -4
  14. wisent/core/cli/diagnose_vectors.py +7 -4
  15. wisent/core/cli/estimate_unified_goodness_time.py +6 -4
  16. wisent/core/cli/generate_pairs_from_task.py +9 -56
  17. wisent/core/cli/geometry_search.py +137 -0
  18. wisent/core/cli/get_activations.py +1 -1
  19. wisent/core/cli/method_optimizer.py +4 -3
  20. wisent/core/cli/modify_weights.py +3 -2
  21. wisent/core/cli/optimize_sample_size.py +1 -1
  22. wisent/core/cli/optimize_steering.py +14 -16
  23. wisent/core/cli/optimize_weights.py +2 -1
  24. wisent/core/cli/preview_pairs.py +203 -0
  25. wisent/core/cli/steering_method_trainer.py +3 -3
  26. wisent/core/cli/tasks.py +19 -76
  27. wisent/core/cli/train_unified_goodness.py +3 -3
  28. wisent/core/contrastive_pairs/diagnostics/control_vectors.py +4 -4
  29. wisent/core/contrastive_pairs/diagnostics/linearity.py +7 -0
  30. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/agentic_search.py +37 -347
  31. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/aider_polyglot.py +113 -136
  32. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/codeforces.py +2 -12
  33. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/coding_benchmarks.py +124 -504
  34. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/faithbench.py +40 -63
  35. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/flames.py +46 -89
  36. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/flores.py +15 -4
  37. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/frames.py +36 -20
  38. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/hallucinations_leaderboard.py +3 -45
  39. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/livemathbench.py +42 -4
  40. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/longform_writing.py +2 -112
  41. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/math500.py +39 -4
  42. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/medium_priority_benchmarks.py +475 -525
  43. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/mercury.py +65 -42
  44. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/olympiadbench.py +2 -12
  45. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/planbench.py +78 -219
  46. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/polymath.py +37 -4
  47. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/recode.py +84 -69
  48. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/refusalbench.py +168 -160
  49. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/simpleqa.py +44 -25
  50. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/tau_bench.py +3 -103
  51. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/toolbench.py +3 -97
  52. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/toolemu.py +48 -182
  53. wisent/core/contrastive_pairs/lm_eval_pairs/lm_extractor_manifest.py +3 -0
  54. wisent/core/contrastive_pairs/lm_eval_pairs/lm_extractor_registry.py +19 -1
  55. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/aclue.py +1 -3
  56. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/acp_bench.py +1 -3
  57. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/acp_bench_hard.py +1 -3
  58. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/advanced.py +2 -4
  59. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/aexams.py +1 -3
  60. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/afrimmlu.py +1 -3
  61. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/afrixnli.py +2 -2
  62. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arabculture.py +1 -3
  63. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arabic.py +1 -3
  64. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arabic_exams.py +1 -3
  65. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arabic_leaderboard_complete.py +1 -3
  66. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arabic_leaderboard_light.py +1 -3
  67. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arabicmmlu.py +1 -3
  68. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/aradice.py +1 -3
  69. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arc.py +1 -3
  70. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arc_challenge.py +1 -2
  71. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arc_easy.py +1 -2
  72. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arithmetic.py +2 -2
  73. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/asdiv.py +2 -2
  74. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/babi.py +36 -2
  75. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/basque_bench.py +1 -3
  76. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/bbq.py +1 -3
  77. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/belebele.py +1 -3
  78. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/benchmarks.py +1 -3
  79. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/bertaqa.py +1 -3
  80. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/bhs.py +1 -3
  81. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/bhtc.py +3 -5
  82. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/blimp.py +1 -3
  83. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/blimp_nl.py +1 -3
  84. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/boolq.py +2 -2
  85. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/c4.py +1 -3
  86. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cabbq.py +1 -3
  87. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/careqa.py +1 -3
  88. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/catalan_bench.py +1 -3
  89. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/catalanqa.py +1 -3
  90. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/catcola.py +1 -3
  91. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cb.py +2 -2
  92. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ceval.py +1 -3
  93. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ceval_valid.py +1 -3
  94. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/chain.py +1 -3
  95. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/chartqa.py +1 -3
  96. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/claim.py +1 -3
  97. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/click.py +1 -3
  98. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cmmlu.py +1 -3
  99. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cnn.py +1 -3
  100. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cocoteros.py +1 -3
  101. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/coedit.py +1 -3
  102. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/commonsense.py +1 -3
  103. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/commonsense_qa.py +1 -3
  104. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/copa.py +2 -2
  105. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/copal_id.py +1 -3
  106. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/coqa.py +3 -4
  107. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/csatqa.py +1 -3
  108. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cycle.py +1 -3
  109. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/darija_bench.py +1 -3
  110. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/darijahellaswag.py +2 -6
  111. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/darijammlu.py +1 -3
  112. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/dbpedia.py +1 -3
  113. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/discrim_eval.py +1 -3
  114. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/doc.py +1 -3
  115. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/drop.py +2 -2
  116. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/epec.py +1 -3
  117. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eq.py +1 -3
  118. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eq_bench.py +1 -3
  119. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eq_bench_ca.py +1 -3
  120. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eq_bench_es.py +1 -3
  121. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/esbbq.py +1 -3
  122. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ethics.py +1 -3
  123. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eus.py +1 -3
  124. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eus_exams.py +1 -3
  125. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eus_proficiency.py +1 -3
  126. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eus_reading.py +1 -3
  127. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eus_trivia.py +1 -3
  128. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/evalita_llm.py +1 -3
  129. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/financial.py +1 -3
  130. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/flan.py +1 -3
  131. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/french_bench.py +1 -3
  132. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/galician_bench.py +1 -3
  133. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/gaokao.py +2 -2
  134. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/glianorex.py +1 -3
  135. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/global_mmlu.py +1 -3
  136. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/global_piqa.py +1 -3
  137. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/gpt3.py +1 -3
  138. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/groundcocoa.py +1 -3
  139. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/haerae.py +1 -3
  140. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/headqa.py +2 -2
  141. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/hellaswag.py +2 -2
  142. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/hendrycks_ethics.py +5 -9
  143. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/hendrycks_math.py +63 -16
  144. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/histoires_morales.py +1 -3
  145. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/hrm8k.py +1 -3
  146. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/humaneval_infilling.py +1 -3
  147. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/icelandic_winogrande.py +1 -3
  148. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/inverse.py +1 -3
  149. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/inverse_scaling.py +1 -3
  150. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ja.py +1 -3
  151. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/japanese_leaderboard.py +1 -3
  152. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/japanese_leaderboard_mc.py +1 -1
  153. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/kmmlu.py +1 -3
  154. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/kobest.py +1 -3
  155. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/kormedmcqa.py +5 -17
  156. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/lambada_cloze.py +1 -3
  157. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/lambada_multilingual.py +1 -3
  158. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/law.py +1 -3
  159. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/leaderboard.py +1 -3
  160. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/lingoly.py +1 -3
  161. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/llama3.py +1 -3
  162. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/lm_syneval.py +1 -3
  163. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/logiqa.py +2 -2
  164. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/logiqa2.py +2 -2
  165. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/longbench.py +1 -3
  166. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/longbenchv2.py +1 -3
  167. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mastermind.py +2 -4
  168. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mc-taco.py +2 -2
  169. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/med_concepts_qa.py +2 -4
  170. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/meddialog.py +1 -3
  171. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/medical.py +1 -3
  172. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/medmcqa.py +1 -3
  173. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/medqa.py +2 -2
  174. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mela.py +2 -2
  175. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/metabench.py +1 -3
  176. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/minerva_math.py +1 -3
  177. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mmlu.py +1 -3
  178. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mmlusr.py +3 -4
  179. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mrpc.py +2 -2
  180. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/multiblimp.py +2 -5
  181. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/multirc.py +2 -2
  182. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mutual.py +2 -2
  183. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/non.py +1 -3
  184. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/noreval.py +1 -3
  185. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/noreval_exact.py +1 -3
  186. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/noreval_gen_exact.py +1 -3
  187. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/noreval_mc.py +4 -8
  188. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/noreval_mc_log_likelihoods.py +4 -8
  189. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/nq_open.py +2 -2
  190. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/okapi_arc_multilingual.py +1 -3
  191. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/okapi_hellaswag_multilingual.py +1 -3
  192. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/okapi_mmlu_multilingual.py +1 -3
  193. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/okapi_truthfulqa_multilingual.py +2 -5
  194. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/olaph.py +1 -3
  195. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/openbookqa.py +2 -2
  196. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/option.py +1 -3
  197. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/parafraseja.py +1 -3
  198. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/parafrases.py +1 -3
  199. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/paws.py +1 -3
  200. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/paws_x.py +1 -3
  201. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/pawsx.py +2 -2
  202. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/persona.py +1 -3
  203. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/phrases.py +1 -3
  204. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/pile.py +1 -3
  205. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/piqa.py +2 -2
  206. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/portuguese_bench.py +1 -3
  207. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/prompt.py +1 -3
  208. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/prost.py +2 -2
  209. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/pubmedqa.py +2 -2
  210. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/qa4mre.py +2 -2
  211. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/qasper.py +2 -2
  212. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/qasper_bool.py +2 -2
  213. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/qnli.py +2 -2
  214. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/qnlieu.py +1 -3
  215. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/qqp.py +2 -2
  216. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/race.py +2 -2
  217. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/random.py +1 -3
  218. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/record.py +2 -2
  219. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/reversed.py +1 -3
  220. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/rte.py +2 -2
  221. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ruler.py +1 -3
  222. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/sciq.py +2 -2
  223. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/score.py +1 -3
  224. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/scrolls.py +1 -3
  225. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/scrolls_mc.py +1 -3
  226. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/self.py +1 -3
  227. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/sglue.py +1 -3
  228. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/sglue_rte.py +2 -1
  229. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/siqa.py +4 -7
  230. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/social_iqa.py +2 -2
  231. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/spanish_bench.py +1 -3
  232. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/storycloze.py +2 -6
  233. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/summarization.py +1 -3
  234. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/super.py +1 -3
  235. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/super_glue.py +1 -3
  236. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/swag.py +2 -2
  237. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/swde.py +1 -3
  238. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/sycophancy.py +1 -3
  239. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/t0.py +1 -3
  240. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/teca.py +1 -3
  241. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinyarc.py +1 -3
  242. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinybenchmarks.py +1 -3
  243. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinygsm8k.py +1 -3
  244. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinyhellaswag.py +1 -3
  245. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinymmlu.py +1 -3
  246. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinytruthfulqa.py +1 -3
  247. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinywinogrande.py +1 -3
  248. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tmmluplus.py +1 -3
  249. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/triviaqa.py +2 -2
  250. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/truthfulqa.py +1 -3
  251. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/truthfulqa_mc1.py +1 -3
  252. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/truthfulqa_mc2.py +1 -3
  253. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/turblimp_core.py +1 -3
  254. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/turkishmmlu.py +1 -3
  255. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/turkishmmlu_mc.py +0 -2
  256. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/unscramble.py +1 -3
  257. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/vaxx.py +2 -2
  258. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/webqs.py +2 -2
  259. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wic.py +3 -4
  260. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/winogrande.py +2 -2
  261. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wmdp.py +1 -3
  262. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wnli.py +2 -2
  263. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wsc.py +2 -2
  264. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wsc273.py +1 -3
  265. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/xcopa.py +1 -3
  266. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/xlsum.py +1 -3
  267. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/xnli.py +2 -2
  268. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/xquad.py +2 -4
  269. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/xstorycloze.py +2 -3
  270. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/xwinograd.py +2 -2
  271. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/zhoblimp.py +1 -3
  272. wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_pairs_generation.py +173 -6
  273. wisent/core/data_loaders/loaders/lm_loader.py +12 -1
  274. wisent/core/geometry_runner.py +995 -0
  275. wisent/core/geometry_search_space.py +237 -0
  276. wisent/core/hyperparameter_optimizer.py +1 -1
  277. wisent/core/main.py +3 -0
  278. wisent/core/models/core/atoms.py +5 -3
  279. wisent/core/models/wisent_model.py +1 -1
  280. wisent/core/optuna/classifier/optuna_classifier_optimizer.py +2 -2
  281. wisent/core/parser_arguments/check_linearity_parser.py +12 -2
  282. wisent/core/parser_arguments/generate_vector_from_synthetic_parser.py +2 -2
  283. wisent/core/parser_arguments/generate_vector_from_task_parser.py +2 -2
  284. wisent/core/parser_arguments/geometry_search_parser.py +61 -0
  285. wisent/core/parser_arguments/main_parser.py +8 -0
  286. wisent/core/parser_arguments/train_unified_goodness_parser.py +2 -2
  287. wisent/core/steering.py +5 -3
  288. wisent/core/steering_methods/methods/hyperplane.py +2 -1
  289. wisent/core/synthetic/generators/nonsense_generator.py +30 -18
  290. wisent/core/trainers/steering_trainer.py +2 -2
  291. wisent/core/utils/device.py +27 -27
  292. wisent/core/utils/layer_combinations.py +70 -0
  293. wisent/examples/__init__.py +1 -0
  294. wisent/examples/scripts/__init__.py +1 -0
  295. wisent/examples/scripts/count_all_benchmarks.py +121 -0
  296. wisent/examples/scripts/discover_directions.py +469 -0
  297. wisent/examples/scripts/extract_benchmark_info.py +71 -0
  298. wisent/examples/scripts/generate_paper_data.py +384 -0
  299. wisent/examples/scripts/intervention_validation.py +626 -0
  300. wisent/examples/scripts/results/test_AraDiCE_ArabicMMLU_lev_evaluation.json +324 -0
  301. wisent/examples/scripts/results/test_AraDiCE_ArabicMMLU_lev_pairs.json +92 -0
  302. wisent/examples/scripts/results/test_aexams_IslamicStudies_evaluation.json +324 -0
  303. wisent/examples/scripts/results/test_aexams_IslamicStudies_pairs.json +92 -0
  304. wisent/examples/scripts/results/test_afrimgsm_pairs.json +92 -0
  305. wisent/examples/scripts/results/test_afrimmlu_evaluation.json +324 -0
  306. wisent/examples/scripts/results/test_afrimmlu_pairs.json +92 -0
  307. wisent/examples/scripts/search_all_short_names.py +31 -0
  308. wisent/examples/scripts/test_all_benchmarks.py +138 -0
  309. wisent/examples/scripts/test_all_benchmarks_new.py +28 -0
  310. wisent/examples/scripts/test_contrastive_pairs_all_supported.py +230 -0
  311. wisent/examples/scripts/test_nonsense_baseline.py +261 -0
  312. wisent/examples/scripts/test_one_benchmark.py +324 -0
  313. wisent/examples/scripts/test_one_coding_benchmark.py +293 -0
  314. wisent/examples/scripts/threshold_analysis.py +434 -0
  315. wisent/examples/scripts/visualization_gallery.py +582 -0
  316. wisent/parameters/lm_eval/broken_in_lm_eval.json +179 -2
  317. wisent/parameters/lm_eval/category_directions.json +137 -0
  318. wisent/parameters/lm_eval/repair_plan.json +282 -0
  319. wisent/parameters/lm_eval/weak_contrastive_pairs.json +38 -0
  320. wisent/parameters/lm_eval/working_benchmarks.json +206 -0
  321. wisent/parameters/lm_eval/working_benchmarks_categorized.json +236 -0
  322. wisent/tests/test_detector_accuracy.py +1 -1
  323. wisent/tests/visualize_geometry.py +1 -1
  324. {wisent-0.7.701.dist-info → wisent-0.7.901.dist-info}/METADATA +1 -1
  325. {wisent-0.7.701.dist-info → wisent-0.7.901.dist-info}/RECORD +329 -295
  326. wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/browsecomp.py +0 -245
  327. {wisent-0.7.701.dist-info → wisent-0.7.901.dist-info}/WHEEL +0 -0
  328. {wisent-0.7.701.dist-info → wisent-0.7.901.dist-info}/entry_points.txt +0 -0
  329. {wisent-0.7.701.dist-info → wisent-0.7.901.dist-info}/licenses/LICENSE +0 -0
  330. {wisent-0.7.701.dist-info → wisent-0.7.901.dist-info}/top_level.txt +0 -0
{wisent-0.7.701.dist-info → wisent-0.7.901.dist-info}/RECORD RENAMED
@@ -1,4 +1,4 @@
1
- wisent/__init__.py,sha256=AG1M4u2B5EcGHVjK-FMR_6hNR3P9q1OzxP3E11LExQY,1229
1
+ wisent/__init__.py,sha256=_tapv1VtSsFzxXjtnrr2XXezxkMMdt4Me1YSKo_fSKc,1229
2
2
  wisent/cli.py,sha256=XKzGIGstr38EowHYpr821c6YuV9Eaw3I1I3NvLztTO0,3960
3
3
  wisent/core/__init__.py,sha256=x1MX4vKpKP3c2FuIHcFly-UkoZwGVnRPbzcFaxr_Jdo,1340
4
4
  wisent/core/autonomous_agent.py,sha256=rtJ5XlWqPys7c1PBFkUBHGg9Ox6_LDjeoiNPEph6NA4,48667
@@ -12,10 +12,12 @@ wisent/core/config_manager.py,sha256=RfB9PT3K1Ns5EsQiSGQ1DSxdlG-cVMdaaUuqQMdCNOY
12
12
  wisent/core/detection_handling.py,sha256=L14US_y4UWaSabeAEzWn6tPM79qQFgwpO-BXvBBO7fA,11183
13
13
  wisent/core/diversity_processors.py,sha256=wm4rFK8qrvYLcDpaJVHe-r6QtDYDRGty-Eyd0nXzcdM,6883
14
14
  wisent/core/download_full_benchmarks.py,sha256=6JVmof4xKC--Lkj8r6v70KGRxPVTqu7YC3nVsb5m42g,60722
15
- wisent/core/hyperparameter_optimizer.py,sha256=_aMhzjUlXGGMQeSsX8gWXWf-zwyTopA-mbbOzPEnbSA,22761
15
+ wisent/core/geometry_runner.py,sha256=KvDfSSob-scJSBB_Wya81rlWpiQohxLPjQxoyQw24rA,35719
16
+ wisent/core/geometry_search_space.py,sha256=JSF-zrwFkJxn33Gie_irTxhZW8M6h7eUpRu0Y_9Ci6k,8401
17
+ wisent/core/hyperparameter_optimizer.py,sha256=_L767LaidpRJN23nV9raWQ0wapclIEPJDX3h72mBQo0,22741
16
18
  wisent/core/layer.py,sha256=on_jCGb0CUhn8dpbY9ZbH3yXk1QoNqR6uUPsI6eX1hU,391
17
19
  wisent/core/lm_eval_harness_ground_truth.py,sha256=46WGnsJLoQ3ALXnXJ0PkTo-N1U2wGDGKyU1kwg-1Ap0,66514
18
- wisent/core/main.py,sha256=dovHLbcRV8zEwKz7nRmryJntBS8QODCzI78Kw9UpBps,4353
20
+ wisent/core/main.py,sha256=oL5ODpKkbM2HR9HLlBUbf2z9JSUkMzeI_8zHeMuJcrE,4503
19
21
  wisent/core/managed_cached_benchmarks.py,sha256=_q26-13t_GLFvkMA3M8tge0Sv6fFJmH_AJqCVl-Dt3w,23499
20
22
  wisent/core/mixed_benchmark_sampler.py,sha256=J1R7Q7LOXE28P2Bydx_NrgblJv35rqdkCPkvZiAxp_c,13715
21
23
  wisent/core/model_persistence.py,sha256=Pr1A3E0MxqwICKsFXmjKVq9KgG8Ip5UbtG-gujGFShc,9870
@@ -23,7 +25,7 @@ wisent/core/multi_steering.py,sha256=7RKYLBBFxEsrSc83KPakAU3WzvVx5XO88tShM2cLKZE
23
25
  wisent/core/parser.py,sha256=v99xCTbiyy2f4JzaVjz7bUtvn-uZEAj4S-u2OPikE24,68987
24
26
  wisent/core/representation.py,sha256=hBl_N9qbr5Gsa7GCQ0nMWRm82RqYEfhd9cyf0PPH5LY,195
25
27
  wisent/core/save_results.py,sha256=NJ7elQzbh3Q_ZefdTV8L-9QyutteWSBWfgTw78OBm1k,13725
26
- wisent/core/steering.py,sha256=oiDIvHVm-n0eCS_Er9kv-Bu0H2JOqns87QhY5Oz-T2A,22790
28
+ wisent/core/steering.py,sha256=LFSEla3sCdGBDprjZ5si-ouFcYdxnflW8pfAgyPEJkY,22993
27
29
  wisent/core/steering_method.py,sha256=Y09jIImhZnRuHkbpeAfJT1hDovXi0bacqt3X3TVSo68,411
28
30
  wisent/core/steering_optimizer.py,sha256=SVMI6xHtMb6xAZYkXggziqjEZ9SSkWCL0PiBj4WMlsc,65353
29
31
  wisent/core/task_interface.py,sha256=4xm6nrduFdXSZpJccahfZWtpI4rZvCQcsS7GCJWpbMU,4906
@@ -34,10 +36,11 @@ wisent/core/universal_subspace.py,sha256=D35PgZE42IwzjdyrTjmPcYou9YKJkbVZPfwCHh2
34
36
  wisent/core/user_model_config.py,sha256=FGzHYo1kD7qDPhpffvDkigV-z7aiLxHAtqiMEPf_sIM,6729
35
37
  wisent/core/wisent.py,sha256=yglyOL-10ptlTVTshgjuXle3PXo-fdWJiZPx8oLyabM,21076
36
38
  wisent/core/activations/__init__.py,sha256=KjKMOPbPMmi1iMQMH6ALkcwa8pZLK8dpbz0cHGVjl_M,1352
37
- wisent/core/activations/activations.py,sha256=htpxdaF18uxknd6ETs_3iBKwWqh3bI3x6JwB-jkfk6E,3175
38
- wisent/core/activations/activations_collector.py,sha256=WPLX0-wVV-HbYTyaXo3okqiiWD72mQ6k5JDC85hpgM4,10214
39
- wisent/core/activations/classifier_inference_strategy.py,sha256=oA8VEbuWKLQPZUvjbXC1_f6jAvLVCII3DHLHvXKJBHo,6910
40
- wisent/core/activations/extraction_strategy.py,sha256=N2m8OWobcOGFDsngan1PA2w6S1zhozwG9bnT_vu8vjo,12432
39
+ wisent/core/activations/activation_cache.py,sha256=Tc-qH26Ht_E9wtQYJHnEMtea3dvB8x9XWaZOAQ2oPtA,13784
40
+ wisent/core/activations/activations.py,sha256=ljDkyCPiFyIt-SA43m-vJh3U2pcYIMenYxm46Vy4B1I,3200
41
+ wisent/core/activations/activations_collector.py,sha256=3aY6VxWVErnWOLdTLY6bapjR9e5kzHOoSV2OaOsDlYE,10536
42
+ wisent/core/activations/classifier_inference_strategy.py,sha256=Nw7xz08vgxJoMAUbCOT-vY118815STjh8iOZGbiBBU8,7054
43
+ wisent/core/activations/extraction_strategy.py,sha256=AkbVZ5TyF3bST4ZOw33HVNVlrbGAb8XrhAxoeDtRPNY,19820
41
44
  wisent/core/activations/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
42
45
  wisent/core/activations/core/atoms.py,sha256=wfPhgSt7NP_4xkyYZA6T62KTM2DUkq1Jj-FJeRd9Kko,5383
43
46
  wisent/core/adapters/__init__.py,sha256=UZkssKYL4-Jl_F29x29MQr6L4899lpK-eFvFa_qGa8Y,712
@@ -70,50 +73,52 @@ wisent/core/classifiers/__init__.py,sha256=AZyqSkAxKUtmNFBHfFfK8DzU2UwelX6Xhl4VN
70
73
  wisent/core/classifiers/classifiers/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
71
74
  wisent/core/classifiers/classifiers/rotator.py,sha256=nc3N-PuZwaBzbtk4n7cetZux1GB0BWPHTV7cACXP23I,5148
72
75
  wisent/core/classifiers/classifiers/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
73
- wisent/core/classifiers/classifiers/core/atoms.py,sha256=iy6oRuPUYJUN8UUpxu_FZ8wc9Yf-ZagFjTC82uZFGqU,26131
76
+ wisent/core/classifiers/classifiers/core/atoms.py,sha256=UrNwIIeJgd7JpibliEas9dinTtIrCuAC1lc1_f_qdiE,26194
74
77
  wisent/core/classifiers/classifiers/models/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
75
78
  wisent/core/classifiers/classifiers/models/logistic.py,sha256=6FNnybW4zZ0HSeFjyEiOQrtNFN4LV-M39mmiiW4buDk,876
76
79
  wisent/core/classifiers/classifiers/models/mlp.py,sha256=f_8Lgh5UZPLLCIKF1b6qeg5MwyFj9tuLrNMUcs-QDOo,1536
77
80
  wisent/core/classifiers/core/__init__.py,sha256=DJVQfGN-oXkaUur5HwH9xPP9-VCBVxjkIt4U_SqriA4,28
78
81
  wisent/core/classifiers/models/__init__.py,sha256=DJVQfGN-oXkaUur5HwH9xPP9-VCBVxjkIt4U_SqriA4,28
79
82
  wisent/core/classifiers/pipeline_steps/__init__.py,sha256=DJVQfGN-oXkaUur5HwH9xPP9-VCBVxjkIt4U_SqriA4,28
80
- wisent/core/cli/__init__.py,sha256=0paqfodiH1pMUIzqLbnc9HB-reukpjfwzmgmfUrOMX4,1929
81
- wisent/core/cli/check_linearity.py,sha256=Ktu59KiXhTEL0U7pkD-H16xOdlhcs33dA08hkHSJXDc,4194
82
- wisent/core/cli/cluster_benchmarks.py,sha256=1k0rOO8jgEop4hWvkJqPNqNmSnmYeQEYBT2QXHaMezo,20163
83
- wisent/core/cli/create_steering_vector.py,sha256=M11P5WkmEiuflR7qNRPA_pPkkZAQcbe7fsg5P4ddirs,16640
83
+ wisent/core/cli/__init__.py,sha256=YBvZ1ko75IFtmmvETRC8wa77wij8HvBO_FSHMYandBU,2009
84
+ wisent/core/cli/check_linearity.py,sha256=weRPt1arFa1a8sSUY8JjyApjLssokMX_kJeiLHDqv1I,5916
85
+ wisent/core/cli/cluster_benchmarks.py,sha256=hPkR0pujWJixb7C4QuqNTZ6H2ElExdTb_FnsDkbE0L8,20056
86
+ wisent/core/cli/create_steering_vector.py,sha256=4iCXFKF-HkxwyUnooOLzH37L2FP6cWOEX8_saD8e914,16695
84
87
  wisent/core/cli/diagnose_pairs.py,sha256=dopxJUgxien33yx6zAhwRVfb_4uvIem7Zpj2lawC1dQ,6775
85
- wisent/core/cli/diagnose_vectors.py,sha256=jWLyk990j7jHtP1XVga6hVaR6LOzF6pzjwhkq--GZfs,17088
86
- wisent/core/cli/estimate_unified_goodness_time.py,sha256=KnT7qWBkKwn7M1iaM62NRQpKdF_3pmoTQq4w7WYbRRU,16510
88
+ wisent/core/cli/diagnose_vectors.py,sha256=cwn9P0Ui-OWBKD5Wl-RmxqR3Ruel_EPUYPg0qiJB-T4,17177
89
+ wisent/core/cli/estimate_unified_goodness_time.py,sha256=dsJLh7a_4iaYoJlsQ_t46CTl_8RzMun2LPJD02jiMts,16789
87
90
  wisent/core/cli/evaluate_refusal.py,sha256=V9MjNEs4XqChnr1o9CrkgFwVVbcrhmNtl-p-0nkZ_jY,8758
88
91
  wisent/core/cli/evaluate_responses.py,sha256=jDsb6qnnqJWivL80DERtvWN5nD7je68FNxVp3TNBBIA,38415
89
92
  wisent/core/cli/generate_humanization_pairs.py,sha256=hspH8kd9u9BnI0Zdh7s2nMvCi45sMIvd9bLSIa9eLio,4200
90
93
  wisent/core/cli/generate_pairs.py,sha256=KT88DaRMv5UcQwkLY9hf8uF3vuq9i3pWSnAL6Nejrzc,7545
91
- wisent/core/cli/generate_pairs_from_task.py,sha256=YppLsiyDJhv4U7aoMBggLshLD5MsRQGirrLrg6oBLLc,4406
94
+ wisent/core/cli/generate_pairs_from_task.py,sha256=OfxO-JpiYncUvRZL6AEbGzHNJm7QhGGZ3d9YxShXq2Q,2130
92
95
  wisent/core/cli/generate_responses.py,sha256=nPEwQ8_EVjRbr0lhsCl8sacQzvt5XgDUOONSwRyI--U,5238
93
96
  wisent/core/cli/generate_vector_from_synthetic.py,sha256=Fc9vcd7-1VTqHy8-gjTuPiCC7FIP1rxTajZf8TFSIi8,8868
94
97
  wisent/core/cli/generate_vector_from_task.py,sha256=O619eALpDrTWMq4Milih_IC3eDViXz46EP6D1A6uduE,10121
95
- wisent/core/cli/get_activations.py,sha256=tXEvtIIWv-98vyched-3MJORdy5HUfelWmE4M8Fx8gI,6612
98
+ wisent/core/cli/geometry_search.py,sha256=DOXjzEspdcnLNhjZKf_Mapi33mMVi2LMXpQNLQkE5zI,5503
99
+ wisent/core/cli/get_activations.py,sha256=dAvlhRIWg9p5Z_9_G7UQmgUB3HjsmBbEADSGwPYGEDA,6592
96
100
  wisent/core/cli/inference_config.py,sha256=iIZa0_ciBbM0Fqat0DJNd-mlWd0ANuYoeZ1v1V4jduc,3148
97
101
  wisent/core/cli/inference_config_cli.py,sha256=CKvD3TpFYDSNtm9HdRFv0uMGInaR6iSHP8aLWYPhrfs,1786
98
- wisent/core/cli/method_optimizer.py,sha256=J17rwW9ZHxXZwWTkkgW7HfOQ_rRynvS9Wd2CqWyuT0I,32791
99
- wisent/core/cli/modify_weights.py,sha256=JhZwUl91lWsZwdiie60FQCxAfcwewtICBXf8XOLlBdw,28562
102
+ wisent/core/cli/method_optimizer.py,sha256=oDfS80biqK8DsuB387KgzehoILebZ2m1_7_aL8eqgpQ,32903
103
+ wisent/core/cli/modify_weights.py,sha256=YHWvuDbrDghIzHCrfCdX5ndpg54ESFgLm6uY7iekhFM,28660
100
104
  wisent/core/cli/multi_steer.py,sha256=GRs0557YabcEwch7P75uWJadHf4Ssa8YE5nTMt8WTvs,4564
101
105
  wisent/core/cli/optimization_cache.py,sha256=nvnmFSy9km3tcquL36bhumCvc6wxl-_Ch3Ju9-MzNOE,9943
102
106
  wisent/core/cli/optimize.py,sha256=4ZpLawTrvU57SUEEhxYiR-Xy5c3mZrE9gSRIxgSRWSg,28090
103
107
  wisent/core/cli/optimize_classification.py,sha256=deg3psSPvERDptRIfv4x88adKL0RM3gCBJGfb0QdWnI,23099
104
- wisent/core/cli/optimize_sample_size.py,sha256=EQpIA-lYkh74x3JKRLlmec0RyVElnC-KwYkqUx1H6kM,16204
105
- wisent/core/cli/optimize_steering.py,sha256=UyBVszZPCjqPu45VdROiksyof2ycauZ0YQ8kRP8EgCk,173041
106
- wisent/core/cli/optimize_weights.py,sha256=2uybF7B7LB28-mhTzb088OeWOeX90H_9CHj1MEWYK9k,54285
107
- wisent/core/cli/steering_method_trainer.py,sha256=UKkur2Ice8zXJwho6TG9G9aO_Lk99G2bcscOUycoVA8,22887
108
+ wisent/core/cli/optimize_sample_size.py,sha256=kn8d8fqu4K1cq8CEW_ufz1iu4-nJh_Y3dl15fCmWCq4,16184
109
+ wisent/core/cli/optimize_steering.py,sha256=-8K_RrojHhInLHUtCgvOVqsozUMuEWZcFuBDDhLoVrw,172781
110
+ wisent/core/cli/optimize_weights.py,sha256=uuiepRuq3xM88rnIxtmlmD7a3ylHS8I6WMsa4psCMoc,54364
111
+ wisent/core/cli/preview_pairs.py,sha256=MJjdYJhXCAxA3BvcRQxQfaVZzoKuj4sQl112maHP5Qk,8056
112
+ wisent/core/cli/steering_method_trainer.py,sha256=AENxXgpP6YlFbelpdq3LsWq8WQKfXY1Ub0L6dic__iw,22827
108
113
  wisent/core/cli/steering_search_space.py,sha256=4iGf7vTHPEE4JW1mQdq_P1vLYWyGlUekoyjMMJgx0Hw,23113
109
- wisent/core/cli/tasks.py,sha256=jyqTOfAQjmuUwS0guRzLjxbmUgLZ0PPbXtgTbAoEORs,43381
110
- wisent/core/cli/train_unified_goodness.py,sha256=rX6rRVYYM66WDvIexiLe-FALBHGuns2kRY94zEsZTmE,29086
114
+ wisent/core/cli/tasks.py,sha256=ckQ3NW8DDrHSpmxm3HZ27e6j9ex_lqKKIMTXQ97Np8k,39717
115
+ wisent/core/cli/train_unified_goodness.py,sha256=fOcw-9yD2tcCwUTdxqxFsfuSsCuEPZ8x7nth1gN9lnE,29079
111
116
  wisent/core/cli/agent/__init__.py,sha256=tWH634YZVFfHmaEEQPgOn1LILIXwdpyJg5nj9tsdFz8,591
112
- wisent/core/cli/agent/apply_steering.py,sha256=HDeX6nhhO87z1BzNdZeaAH-1B-K4twvRNIkoQEKlQRE,6729
117
+ wisent/core/cli/agent/apply_steering.py,sha256=hLjsA7B9Y7t1AHvrs1QhDZm3v_Tqo4NYMCCtW2Yo66c,6635
113
118
  wisent/core/cli/agent/evaluate_response.py,sha256=Xk-nKSeGxTqMX7IF6tjYdTUkc_EITVIUg8XEYqZBSnk,4522
114
119
  wisent/core/cli/agent/generate_synthetic_pairs.py,sha256=QnMuJvunEB3Icmx9TrFeMklh1fjpTMLeu8mkV_dx46w,4723
115
120
  wisent/core/cli/agent/main.py,sha256=aZ7SIuzTmYkFTaygjTf76Fgho9CiRMIOooes5Vd7ueI,5457
116
- wisent/core/cli/agent/train_classifier.py,sha256=zWqZkVb4wZQnJfUXa8Jvk9C-Xbeo8VOU-kCtH4TVs_c,6545
121
+ wisent/core/cli/agent/train_classifier.py,sha256=qhJlHy0AKdmWJsVg_ZZwLjNL-G8rd6_oSfWtNyx85R0,6961
117
122
  wisent/core/contrastive_pairs/__init__.py,sha256=AbaAf-t_nyVVy_vLjp8WAlMDmNun3KNp_GMWAK25r9g,429
118
123
  wisent/core/contrastive_pairs/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
119
124
  wisent/core/contrastive_pairs/core/atoms.py,sha256=_zghw6c8iisW_SqBIUCoAnzhc5q7t5EgZ4zzTPxeLwQ,1129
@@ -125,11 +130,11 @@ wisent/core/contrastive_pairs/core/set.py,sha256=sVJ2zvlVaFpMPgr4dYe-FMhT6sPS0KO
125
130
  wisent/core/contrastive_pairs/diagnostics/__init__.py,sha256=X17XJthKEO4gUxUqBR01ysPwf7jx4LZoPKu3QZPK8Xk,2366
126
131
  wisent/core/contrastive_pairs/diagnostics/activations.py,sha256=TUNxU-HV3oeQxFbZhKo-OISM4mzR-Bib0naHmbqWzk8,1736
127
132
  wisent/core/contrastive_pairs/diagnostics/base.py,sha256=uBi8PdTd6BRyy0lmGjAZLTZdgiiWwPNtsmKkBFCmlD0,2658
128
- wisent/core/contrastive_pairs/diagnostics/control_vectors.py,sha256=C3QCDHDgrL7LJxW40qMXGtEptmLuZ-oL1etclNmOIrU,120062
133
+ wisent/core/contrastive_pairs/diagnostics/control_vectors.py,sha256=BAUOtbq1mAIEdXPWtiPUtfq_NggekteByvfAy5F7ilY,120153
129
134
  wisent/core/contrastive_pairs/diagnostics/coverage.py,sha256=MpT6_IdLqtMpav6mOCiNuemBVFvxWzkUbj7j3ZNx-48,2761
130
135
  wisent/core/contrastive_pairs/diagnostics/divergence.py,sha256=Io3AcGluJogz4qENWu0ivQyFR_5bLN49BzCTI7DIVa4,3430
131
136
  wisent/core/contrastive_pairs/diagnostics/duplicates.py,sha256=s5JPUdWcndX_kWAwP5UmXxEYNkW9DAYqtKwntIRoAiI,4755
132
- wisent/core/contrastive_pairs/diagnostics/linearity.py,sha256=ElTjTYub0txZamCf1IPunYShTCv1VdfHdQqGhPse4NA,10842
137
+ wisent/core/contrastive_pairs/diagnostics/linearity.py,sha256=j4-YvzQbysV2dC_mX7mBhwHqj9pmJ6IWUf1cul3pPmY,11169
133
138
  wisent/core/contrastive_pairs/diagnostics/vector_quality.py,sha256=EmgH05y5j9EbdiNxCc3FuhIvCM1DoUsGCgQVMkYSjYk,22311
134
139
  wisent/core/contrastive_pairs/huggingface_pairs/__init__.py,sha256=PYur_GOluHmwaYUS6rxsoas5-AmrG7OhVGoPteuJB-c,76
135
140
  wisent/core/contrastive_pairs/huggingface_pairs/atoms.py,sha256=iIGJRHpDKLskBE_nNOVk9R5329OooktQ9r92_DZNq9A,8826
@@ -138,8 +143,8 @@ wisent/core/contrastive_pairs/huggingface_pairs/hf_extractor_registry.py,sha256=
138
143
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/__init__.py,sha256=HtFG8o1n69SkHgbkjqPXXqsPMPlRUzFu6UWqQ76c268,2522
139
144
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/agentbench.py,sha256=tdJznfmBU0PxtHwzgXVIJy5qRCGPQ898i_-Mu2FFoF8,7936
140
145
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/agentharm.py,sha256=hWEPf8mKIgZDZuisG0YMtAr04c33lfXjkRT8-wr2zfo,10504
141
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/agentic_search.py,sha256=fKSzaCY5Y2mRjRYr2OpP8RDQ6GN6LU3DHepq3rlNV30,17563
142
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/aider_polyglot.py,sha256=u1KoeWTGdvO4jcxde56Jfrn_52O3JQPgxnThHgHTJ18,7335
146
+ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/agentic_search.py,sha256=OSKPksEL5gs_SLykD7_PIrBV6iLCPEkyCkWxlJsxdPw,4249
147
+ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/aider_polyglot.py,sha256=U7_a3slnr6xyPTmHIC-DS9kv6sJaEigCwEpp8o0wRZY,7229
143
148
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/aime.py,sha256=v0xmp650bG5NYHynaookiXrxstdZOU6zwYOhTtguPMI,3584
144
149
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/aime2024.py,sha256=iWtVKnwr1dS4k-zDU9nS4qsN3hnoFNjKb4JOIHsRuqM,2234
145
150
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/aime2025.py,sha256=l2NC4sA7yCpsEx-NtyQmj7x19KfXHcprsYhxVA8a6OQ,2132
@@ -153,14 +158,13 @@ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/basqueglue.py
153
158
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/bec2016eu.py,sha256=jRg17Bd08mKGsbtC7xt6-5EfLEr93AUfk4578wHebS0,3806
154
159
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/bfcl.py,sha256=LZmvmoJFGU9LUc5xxHbugjWiD__rADJkNrmIHJTC-nE,10536
155
160
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/bhtc_v2.py,sha256=aVRRHAkLIHxpbSqYFXPwIYKdQLTo6X05_Rg3Dof4q_k,3080
156
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/browsecomp.py,sha256=g0Up4jGNzVf0g8SQtBdnCoZvNLTjLe52OV4hwOa7uto,9919
157
161
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/chain_of_thought.py,sha256=QsaJFZiFauHzlenUon7P0FnB9cDdMEVGDpJK2nm_w3g,3160
158
162
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/chinese_simpleqa.py,sha256=2q7r_PplOvbYUcNzHsk3JywEb5j6QjD9mR-Fyf7xw10,7107
159
163
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/cluewsc.py,sha256=yhi43rtNT_zjfZ2yg2BuxHI0kSDSETKc-tPk9eeFqgc,6300
160
164
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/cnn_dailymail.py,sha256=NSWDMh2ScUsvC_WVAgIWfGkkY64F_lPbbj9U4FokD7s,3200
161
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/codeforces.py,sha256=TI5tRScQ4jqC_h8mV01fyZOddyR6Y4J0Vb8USBOB9yY,11774
165
+ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/codeforces.py,sha256=stimcsMEhPgJJfhi9PSXftaAhY5jiSEQGKpFP2yZPv8,11300
162
166
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/codexglue.py,sha256=voaKHYSQV91mkY9pqHehNFqO-nnAKxuq5T0r4JRVsm4,3352
163
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/coding_benchmarks.py,sha256=QXJBlUj_rJIwXqn1uNouyT76hBgwOHMYJabaLIfE9ng,25867
167
+ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/coding_benchmarks.py,sha256=VjvwwWGsBfCIMPm1uRQGwEjRaeCjp80wrUUP8-cwo-U,15417
164
168
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/coedit_gec.py,sha256=a-0kJQFcAmoLNFE-dV-G07th9qXPQE0IuAXPcZaVEm8,2605
165
169
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/conala.py,sha256=auULgkRhG77vy2rAHJlpRhAsdTLrIZuq_XKK9PZF9tU,4075
166
170
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/concode.py,sha256=qhrgzJsJUnCG_hg4ouw_rMK5VFsecTwa4q7h_kXqMK8,3400
@@ -174,16 +178,16 @@ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/ethos_binary.
174
178
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/evalita_mp.py,sha256=Ioy59ZiNVnkwHV7IR0A6wJI3vFASr8o6xcXdrHjOT3I,7719
175
179
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/evalita_sp_sum_task_fp_small_p1.py,sha256=N-LunVyfI4jClY5svXq088V7N0QZR3aJqEPCBwmPjOg,3078
176
180
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/facts_grounding.py,sha256=Oyf65Wn2DzE4tSwvMkR6A4XAQLmmKqBNLvLuTgRApv0,6506
177
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/faithbench.py,sha256=qX_YGB9MBPfs_NhG7zfkxNTSiU-SFPbHfT1C01vkpMA,12124
181
+ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/faithbench.py,sha256=MxqTqaFio6oXWr3YwA2EYitii3KF8zdtbK-s9uYrgIQ,10220
178
182
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/financial_tweets.py,sha256=vUE79yRr5oE1vjFZ_wL4s6XFw7UuakWwrGf5UF24iAw,3635
179
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/flames.py,sha256=wu-LUpq9irabMqMXfCHkfJui2Uyvj7LM0mqq_Jp_2yk,11430
183
+ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/flames.py,sha256=UiKaVxgjpv_QhdMf1FTnPIr_Ebe59YHL_goN5hnSSrc,9034
180
184
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/flan_held_in.py,sha256=Xzx2V6j8ihoYwC4SYG226SXXEH5j9G2vJFOMT0Y7pR8,3650
181
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/flores.py,sha256=MwEv7rWS1_e9PMdmC88SM7yNkh189ZJgbhQXKMvmjT0,23149
182
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/frames.py,sha256=q2BY-zNDQ8pK_faY3JUxA7OUTfMX_kbOkILg-tjaYyc,5009
185
+ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/flores.py,sha256=03d8WzoOknwOj4pssiNDCrUdWXKayp-2Wj1McfzkUbA,23418
186
+ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/frames.py,sha256=9JZ2LCbIm6yASe-w_sWo3XpgWpnjn7GXgQzY99F9NPc,5601
183
187
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/get_negative_example_livecodebench.py,sha256=lrhwIQtQupD-xM9ypOamMcPuohkYHIb5uoDvmvgF8JA,5129
184
188
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/get_positive_example_livecodebench.py,sha256=k_PSaWtokvvBfysz3goFh-ZneXGSAKDaLQ22kkMl5oQ,4622
185
189
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/gpt3_translation_benchmarks.py,sha256=aOsJ-l67Noc1qnq_orp7-iw8wSNLX-VCa9C_gXX7MAc,3773
186
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/hallucinations_leaderboard.py,sha256=ie_z2eq8XkY_ZzedZBMAoMgD0s1-TYUrwlNmpae7MOI,15385
190
+ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/hallucinations_leaderboard.py,sha256=IaXs9oFXkx7o8PRnjrMvYHLmh-npyIDAfoLZ8eKWGSk,12412
187
191
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/halueval.py,sha256=N-nwmLP_Zsni0IvjgB2efPSGwJ0m95oW20eyx9Um0-I,8833
188
192
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/harmbench.py,sha256=HyHZYRFUSZFgJ7atINICsCCK3t01SofCaqbzUjdqrE0,9575
189
193
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/healthbench.py,sha256=_APNLxYsMgu3rJigXJwWdB83qL-X3pBMRJF0fzAPWNo,6831
@@ -199,16 +203,16 @@ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/ledgar.py,sha
199
203
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/livecodebench.py,sha256=gQjkxgF7qByJLOHN2JQwOgwY4IkJUWw7kESNyQlkT1g,1964
200
204
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/livecodebench_contrastive_pair_generator.py,sha256=btT2P6utHmbejctjRelREYO_JCqV_Rhm4BWFJix11EQ,16808
201
205
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/livecodebench_v6.py,sha256=4S6wv2yt6-s1mJeme-iEmTgXFHfOjBmpL4e6Y3QiseA,9486
202
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/livemathbench.py,sha256=ugdpVXr7IP1S3YYsqzA0OQaXsYSW82OreUTMcK_QwhI,7792
206
+ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/livemathbench.py,sha256=G9G9ry5QVStYsZTFPxukH1xLanRXm5n68pMJVxRAFEk,9106
203
207
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/llama.py,sha256=0GP8duCOy1u-G--s4VgIWNIihH06-0xSiE9GCqK82UA,3722
204
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/longform_writing.py,sha256=q44Q7zyqJUuWrKek70YXR34BPW5a5z9mojf5C1SIxxM,12728
208
+ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/longform_writing.py,sha256=NskHJKjm9NBTisNQXiYh6doLVDFl4G92JJY_QEjLXhQ,5982
205
209
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/m_mmlu.py,sha256=RDZmJAZfJdpxXwrbhDPFBOClNDV2gMf9TaBWNhVF3Gw,3724
206
210
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/math.py,sha256=1HUMRBIN9b26aWOcyM70tEBbLfzIiZ9uw3OmKZFBpEA,5980
207
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/math500.py,sha256=S5OlCvbvBkGN6-m8YvNXKZTSMnRM8I1bNF6MtnDSnYo,4760
211
+ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/math500.py,sha256=h3oA5S9Kuq6FLTVYpGrhDemjh8zyGcZLnx2NMDC_9OM,6039
208
212
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/meddialog.py,sha256=n-Mh0iWhd-v6avJVAIKe3juHn_D5CrqsXCgERQNK8uA,2616
209
213
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/medical_abstracts.py,sha256=TP1jcbcVRaDuwwSI_CusIwlchI2QWfMhsCa3rHLH_RY,3576
210
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/medium_priority_benchmarks.py,sha256=-NHzHBPM8I0zVUkdCO4gsN0RyZ3vKJQHyP9RUNfKOso,30169
211
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/mercury.py,sha256=_9EotHw83vgaUDxCnb0eS8TI5_hik68jlq9HdZy6lq4,3518
214
+ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/medium_priority_benchmarks.py,sha256=VMjoaUC6WSFo86BGJa8uYStfAC839uoE0M-3lAmKa_I,27316
215
+ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/mercury.py,sha256=zO81ogmkR-_6r_wZ9asLg3NUY3emeuzEfTF0YVT5iYY,4327
212
216
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/mmlu_redux.py,sha256=VX9mKGZZtsaVggyXxfFKOL2pIs-V6Ah1UXN9rktsBCA,7611
213
217
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/mmlusr.py,sha256=Ib5IPMeTp9m-Zc90YlxSXxJiZWo3tw8WasbuUL33yG0,3974
214
218
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/multimedqa.py,sha256=0ijOZPjd86pZlUqDGHKK-QRJcJSSWPvbUglNy0uSw7o,3811
@@ -222,20 +226,20 @@ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/multiple_js.p
222
226
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/multiple_py.py,sha256=TjowpFLoraDyppMPa7375Ol3NwGs8H1_E1xa68LAhxo,406
223
227
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/multiple_rs.py,sha256=7GtAWK7iig8vaCW-lI77HXOQCSnuXiN8q16XuvrIyRo,4351
224
228
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/non_greedy_robustness_agieval_aqua_rat.py,sha256=zcyMAZv7LDOgNdsU-gxVX5PaL8W6IYXOFeaP_l4ZqFw,3461
225
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/olympiadbench.py,sha256=moW_Y_4Tqgnr8bsFtWXmjkFmEUjjjtmngD-caiqdA_4,9644
229
+ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/olympiadbench.py,sha256=OrJQ55-YwI2n3m0wOmKfwj-h-IE79IZIZPLKdJzamKU,9185
226
230
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/openllm.py,sha256=OLl__FMCKuOIcAq7-nHif4uinknDFQjFKy1yUFaDvhU,3796
227
231
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/option_order_robustness_agieval_aqua_rat.py,sha256=GOGncOk_KiYVVjNXCgUaaVgxiowRF8avM8PWpaUng_o,3479
228
232
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/or_bench.py,sha256=R_tpmbnxXHAUEb47CA3s94F0LcjinO6mB_WMZrl2I6M,11101
229
233
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/penn_treebank.py,sha256=hp1ct35XCY4KQLtJOA36t7K5fl72o9CHDF374oJzBS8,2601
230
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/planbench.py,sha256=c0dhZOub2H9knEGC0eUjDO6dZJ8Ss1R82kT1Jms_VNs,11809
231
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/polymath.py,sha256=D69JfFd7NPoJAQ-AUbuvqvWMHwLS1TJMhoPrXqeNY3g,13058
234
+ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/planbench.py,sha256=CWSq7e3F_NCeiINLkksQ7truDYmuC1YkVgOgUvVuqxU,6250
235
+ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/polymath.py,sha256=lI0Fepaf053ygAYCTnbwoMLCSAlWipctJc1uibMog3A,14194
232
236
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/prompt_robustness_agieval_aqua_rat.py,sha256=g-h6DrOIe4zdNfDUUAqoE90dgBUFbIiEZcxH2Aq-Z9k,3427
233
237
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/pythia.py,sha256=7cTgI1og4b6qSP3LCkG6m9FRnkmP1IzjT3O7ltis7II,3791
234
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/recode.py,sha256=WDeylMHDuZWoEMcLO_bubLxCJBR2X0HaXC8B9ri2jwo,4536
235
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/refusalbench.py,sha256=My4ruNBb8QeWArfkzt7cD9eglSAEiPLjPMtP-tyALUA,11940
238
+ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/recode.py,sha256=V9n8nwSW4Tz3L0bXcZWqQWpin7PJWuecvWqaYesrkVA,4867
239
+ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/refusalbench.py,sha256=CPORdQECm2GHTwzYfwYcdY0EfmMQXSL6RwPwQeN7ccU,11767
236
240
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/scicode.py,sha256=RQUEzLatM8Dys-CWsMahKWB_eAILKTS8exc99REG0Qg,9495
237
241
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/self_consistency.py,sha256=QclAgUIwl0YPTBKHtlOyI1gSzdHeATmL5-kxgB16tPA,3035
238
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/simpleqa.py,sha256=mRsR-tBOfJXChmhSfrbX3J6XWF0P0l3iKR88mK6ov2I,5008
242
+ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/simpleqa.py,sha256=20ptnbwjpli4wsYiQUZCdynImpcdcSw4gHxG64Lgeqw,5712
239
243
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/sorry_bench.py,sha256=orRjSL-mfjtXBWpwd18yNo-XVdUaKTK0RpWLAuiKLhw,8348
240
244
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/stsb.py,sha256=PB7T85YTTN1W6LvFa61ZK2OI02NZqdOsGi0AyFew2dY,2544
241
245
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/super_glue_lm_eval_v1.py,sha256=5sjnFalFsICF29k3DIyJ-PGM18yCA8IBQNsJAuZMWBc,3858
@@ -247,10 +251,10 @@ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/swe_bench_ver
247
251
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/sycophancy_eval.py,sha256=qAkLC9LxYQIyKc-k-v1pFdHcDzfT7yLhDpX-9l6eIhU,7656
248
252
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/t0_eval.py,sha256=1eSUGHSbczaaAYWov4WhwgV0p2bHP-B9BOMUdhIXin0,2577
249
253
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/tag.py,sha256=IBphPJy8Dz0FgEF0WIDMPdcErMiMnX_UbwEBF-hmffg,3572
250
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/tau_bench.py,sha256=rzUmgixtJGa3WyuTt_TsJyQdiYL1nQ6aCVKhMYKzbMk,11306
254
+ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/tau_bench.py,sha256=2tkAQhh8-SiiN0U6Kb-MJTJsrUfvJEGZ4UWpXqcqMF0,7210
251
255
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/tmlu.py,sha256=E2_2XoupBm_LdSiX5Wi6BYabHgXabItq4r8lKh-c-Mk,4550
252
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/toolbench.py,sha256=40my4F9l1sIAgwXJPFWjRIBx05-CotsH6idQnDIZQ_U,14102
253
- wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/toolemu.py,sha256=91KVi6pB4P_4GR9LKeCAqOFUWiyTPtgDh4YPYlWGyzE,13678
256
+ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/toolbench.py,sha256=pWEhmL6qGdir27hJk4xhyTpk3Ei_25txe0yfdW4CZVM,9480
257
+ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/toolemu.py,sha256=QD2BwdIL21l8Wd_AlKMcso-NT8rAmePJK2H2-qAqh0c,8635
254
258
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/travelplanner.py,sha256=x2_P06ZEQFT2tRNtii3NzWXCH9ItQVZG8t4GN_vRWYM,10133
255
259
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/truthfulqa_generation.py,sha256=TW1T3_VVfZ3RAQ73xKojUb2sI9eoYMbEEyBUKdqtFOg,4591
256
260
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/unfair_tos.py,sha256=UnZFX-O-nTbNn3Cic7kZ-x0msQCxC_CEMfI5Ft2OR_c,2805
@@ -268,9 +272,9 @@ wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/wmt_ro_en_t5_
268
272
  wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/xsum.py,sha256=zzIp0MeqI7XccHlNEaqL0w3OR3Ukt6lqMYVZQmksAMs,2907
269
273
  wisent/core/contrastive_pairs/lm_eval_pairs/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
270
274
  wisent/core/contrastive_pairs/lm_eval_pairs/atoms.py,sha256=u8RTCYlpUX7X3DHLx9aPlid3xrYPQK6f7Kymj0ps7Ug,9056
271
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_extractor_manifest.py,sha256=GUvxCEOKl06gzjl_UTTOsCWUfTXs5MiB1IsQQukY_e8,30200
272
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_extractor_registry.py,sha256=Wq5_AOWRUO2m387YaLmzWPWfMmwT96FmaAz0N0_JV6A,5255
273
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_pairs_generation.py,sha256=x5dWdceSnMetwmxliUFmErAi0DHXknSBUQ3SSXBznh4,2090
275
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_extractor_manifest.py,sha256=DcqdQeilSKjY5UYHA-qHFljwjdrfM-1PvaN0CRsdf18,30408
276
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_extractor_registry.py,sha256=N8VpOE8JG4y8GQM6sxjzqI8M0U06KdjzIjrxVwxYS7U,6005
277
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_pairs_generation.py,sha256=mCaFhxXoE6EiknkvS6BeznBqOLfeLbmWKULRPuyht7s,8425
274
278
  wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/__init__.py,sha256=yxkFjHDEpD33FH1-shObI-jsperBSADjTqZxQrzC3QA,16627
275
279
  wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/aclue.py,sha256=jreFoIxdiaOZM-qsv_L8dAkJxXm21ajQR0XUY8LJP-k,1321
276
280
  wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/acp.py,sha256=2txurLRPPG4xBLMd0a8LJjF4XAyQHmQon_gp4PbTl4g,1902
@@ -425,14 +429,14 @@ wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/xquad.py,sha256
425
429
  wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/xstorycloze.py,sha256=VLiKWLfyDPJnk_L_aLchZVV82KCIU9T6IgSvzYS_ZBA,1186
426
430
  wisent/core/contrastive_pairs/lm_eval_pairs/group_task_manifests/xwinograd.py,sha256=RAkchOuJtal6YXe7U1kiNtLJOzepczRuRNqzv3CNKm8,636
427
431
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/__init__.py,sha256=qjtVqnPSAZjOg-ME9QjeOEkgJR4rImyfGImxEwAEfmQ,7278
428
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/aclue.py,sha256=uGbZtgDttvJE5BeAqpkUWOEoUVVDo7APzFXXjAh-HQg,6783
429
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/acp_bench.py,sha256=gX67-4ZALgLegw7rcKSuWfExERxPqiC-1ZAX2rahjaM,8489
430
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/acp_bench_hard.py,sha256=SyIMnyCm5BYWVZmA0x2KzO8otv05e5FQ-vq9mM7u6ec,7582
431
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/advanced.py,sha256=KFexdT3Cmu-cVGzYruT_vHIHJq4-cdhDgii4UEN_cPI,5298
432
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/aexams.py,sha256=3xIlyGLP9TKhBo3Z0OmLY1a7bQsshnHgXFTxkt1RKrc,7275
432
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/aclue.py,sha256=ugSEM2fAh1pYaABSjCKNIjTLxdbAGMpw9-ktZFEIGwg,6685
433
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/acp_bench.py,sha256=Qo7dHcSVzqcNLOkPCfwX8MLQ-XEn2qneQLV_ELdBVmA,8391
434
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/acp_bench_hard.py,sha256=qgrCO7FZEDzHMJB5OhT7-DVTTQIFk1ztlr18fKrl52Y,7484
435
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/advanced.py,sha256=XnAnsrWEfz8TyiTJqWN1wgTHtK8oYwX5EOR0SXkRNcI,5190
436
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/aexams.py,sha256=CwQ7aee2fWqOkDOD4Rj904cZcxE8eoe-4Zmgpv7aRlg,7177
433
437
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/afrimgsm.py,sha256=ozlEmQHqxsaiH5jb1D1WI9aToAk6psQkduDwq737F5c,3527
434
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/afrimmlu.py,sha256=-hH6AKLfE7k09FP4Zt_FFeapGp52BSpYaduyFVhDzKk,4139
435
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/afrixnli.py,sha256=aYpA28yQtnwMfstzyZJDakPhV18x84fVJDIKMZ_d0aw,4466
438
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/afrimmlu.py,sha256=1iyuSG1FMUYur7MO6G7PL7XwZZelGWFixTJ_tE2KC-g,4041
439
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/afrixnli.py,sha256=SSSrkGOqA7paEq6P7NAWVIzwpMZMs5iEPtL2Uc6tQZc,4449
436
440
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/afrobench_cot.py,sha256=XXg_BHyIHOt_NezgVJ5K8aplTR_rTiX1Eg154L3sTI0,3319
437
441
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/afrobench_mc.py,sha256=1aRM-_bxDOqWWc5WCfaes_WIbgfs-zurYQkP_PxR660,4114
438
442
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ag.py,sha256=4lDiGXPZHhThl6fa6DKho14cGEWLnKDPP_5fdTO8gUg,5438
@@ -441,311 +445,311 @@ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ai2_arc.py,sha256
441
445
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/anagrams1.py,sha256=VMEhdDEjXyQscQGSXz98H42LZ75XTuS-QcrARzWx-QQ,3141
442
446
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/anagrams2.py,sha256=gnfUcTl26pUmoYPGDg5SUBasrmCK1ggB_kIQYsijNeU,3141
443
447
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/anli.py,sha256=YunlhuOmXsEO0ic8AzTgOz5I7LaYib1IRdTwFTllErg,4808
444
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arabculture.py,sha256=_tSYyWpXsmFOMe1dR4DLY9cG49la-QC_2J0s7XG572k,7337
445
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arabic.py,sha256=8UEvnJ_q1orPqcxOlKCmbwVKrXK33fmimLwa2YvDsGM,3766
446
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arabic_exams.py,sha256=Kxe471_ial8PtoJD607yNBxijquSVQMv24DMFdAt3T4,3953
447
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arabic_leaderboard_complete.py,sha256=1HbKhGtw6YL42WRTjGHZbLxUMj5Hp26gFrHM-2Vhmzs,6893
448
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arabic_leaderboard_light.py,sha256=cicxXygeMxPteGOECVtraOzvdZRrbd1YW6lw4stvXAA,6863
449
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arabicmmlu.py,sha256=9JM3zPOImCsegFxxVLBkd-QwH4JylErdDTZZ_vOx2s4,6754
450
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/aradice.py,sha256=m7VYmrliIxYsRSnmk7ySVaMtx8sS10esNAeeGC6BYSs,12083
451
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arc.py,sha256=cJTEsh7gR6xXz5of_Al_sVxduGm--tXAkjBS_VRxZng,5238
452
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arc_challenge.py,sha256=TWSj3EjOBZy6BPxFDwxdyplzOQCwtjspZHuIEbG6xng,4121
453
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arc_easy.py,sha256=uJc-wlODwrMcrD6Edn1ZJMIJaZfNMA-thPe6jOMFnWo,4076
448
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arabculture.py,sha256=7tByFxwt6NmXlnOe2MtdAhktfDivwmH41l9R_0i4lnM,7239
449
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arabic.py,sha256=Z3OC8h65GDsx0AX02Oc-YWaV4pPr7KDe9NS7NXlIipU,3668
450
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arabic_exams.py,sha256=YYIenYtFxABAp8GMLoDYovFWG4Gal6XZXdk95yhznCY,3855
451
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arabic_leaderboard_complete.py,sha256=NIMsdln4vJI3MLeuheC8NigPePqFmkTW1hFzNzOxmaM,6795
452
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arabic_leaderboard_light.py,sha256=jj5HRRyOp3dVl0KNDiCUK2tfxTzJYmyWc1_04pbb2xc,6765
453
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arabicmmlu.py,sha256=S9O5FfgOUYtMf-WXnB1RnREE7-svYoyg-lz24ItWnYA,6656
454
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/aradice.py,sha256=0GuCPEHvO1jM0Wur3PQg-syczIvCWvBSjsWpsMvIlb0,11985
455
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arc.py,sha256=H7MJ0J-sjQC_odavk9yNNmaJpkjm-dZpt1jKW2dpHEU,5140
456
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arc_challenge.py,sha256=pLZG8oZJ-Rmc2midYKwFCL6X7kfejszEcZRlVKIaPf0,4024
457
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arc_easy.py,sha256=L7nRJaojsxhKA9q8YJr1dqvcWepuzBt_4bz0KQXLtMY,3979
454
458
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arc_gen.py,sha256=Eokj_PknnpqUK264I68i0DiAjQVcCQxH5lhDPIZ_zes,3881
455
459
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arc_mc.py,sha256=TXvAaLY4wAsHuozHBJ4Dp7ryOG9bA4B_AyhYzS-8WRY,4121
456
460
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/argument.py,sha256=GJt9XdLoXcnJ7_Obat9nK-CenPZNvIaFva7e5XeeXf4,5494
457
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arithmetic.py,sha256=rz6kaUDq-4HgW8ERh48fFWoHfBQZ-jClYD5YB1PIZa4,3880
458
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/asdiv.py,sha256=8xtQLBuSM31XXhhDH_PvsWH-D4pQrnY9Q7plGafjdF8,4215
461
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/arithmetic.py,sha256=38DtqaahEE--j8ZgSPaJzrCYf3FyWNlWm2Eo52Pmw_M,3826
462
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/asdiv.py,sha256=NIJyDHcKGAUYjp68MVDIlp7WVBJtfKQkMsrDvT9HyKU,4161
459
463
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/assin.py,sha256=w0CEpwyxBfHa5SDXEUc-rTbogw9f6vqT7gPiND-XNa0,4158
460
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/babi.py,sha256=RkooITV_aXF58uK69v3Ohj15bPOJS5yYTzuDtLZfNSE,4020
461
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/basque_bench.py,sha256=zZEJMU0SroZMWxI3dkFrlk5CdpZeRJqLbzwxDdGUWQY,6065
464
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/babi.py,sha256=VrYfjZXZBk9RuDdqytZKMV9lnvw9GbWc9GkHdF-Ux6w,6105
465
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/basque_bench.py,sha256=89rrlIBniuODiwhU4r0y2kIZ4RudOH5pTgSfHWU4tTQ,5967
462
466
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/basque_bench_gen.py,sha256=Er8ufToEqS5vRAbl5eHTz7Ihwrv7WYCHdm_WN-GOqWI,5983
463
467
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/basque_bench_mc.py,sha256=F_dnN_Yips5nEX9ZW-nOM4HvA5a6XQ4-WCkeUsig_Ac,5383
464
468
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/bbh.py,sha256=QyL31K5p1pSAoOeaThMXcqjXvn5ARZeyn458y6wIpVw,4730
465
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/bbq.py,sha256=VRMCL7kVxeQidjiQpkMvnJeq6sUyrzGrerLjaLcHi88,6730
466
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/belebele.py,sha256=ksyM3kwlAXoATQxNel6FFSLMyDvxX2uxmevUDUzBCBM,7253
467
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/benchmarks.py,sha256=uNsdQoa3vnF7A4E-tixhO9RpCiYv0t5BCroo4wgHMRg,6052
468
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/bertaqa.py,sha256=qX248uM7Urn88jpwh1ALnAXRxjV3QDGDnGzb-sLL854,6561
469
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/bhs.py,sha256=41eIftkv4auWeZjROUfu377zPcd1Q4_yBRNdDcau6Lk,5982
470
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/bhtc.py,sha256=NX5M5V7X0DBBSlEerITZXlc_Za3F7uOc1VwtDuX-TwM,5816
469
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/bbq.py,sha256=zprXP8p3zKRTPtGWjPO-GVcrpdf6nzrGssonw9AgZlM,6632
470
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/belebele.py,sha256=yID74I8e0QJIBHS5XfMpaaRQ6B88P8uhqUTvtsKZcv4,7155
471
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/benchmarks.py,sha256=5qTiZKdLlpJ2NInL3JrtbeevVC5Vdf8s2vgV3wg-WYw,5954
472
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/bertaqa.py,sha256=gENly4KQM11xUc7JplTO9tpEsecDwGx7EI8saNuDCoI,6463
473
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/bhs.py,sha256=CEVXu-cfKba-SArF0Z3-48UzLjIkAZa4P9we0dII-Dg,5884
474
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/bhtc.py,sha256=VzOGeGDLyJhhDfJiY29Kl1Fp0gFEGU2vVTpk7mLJJU4,5666
471
475
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/bigbench.py,sha256=Yww82fQuUbEANLotw7vP7m_EDoA9CNRWPvgrB6B-wvM,6613
472
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/blimp.py,sha256=t0qiCdHfIxyHN7gaKd1_-Jvctp7yEHhzR_AkBmphP70,6754
473
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/blimp_nl.py,sha256=rGQGExRVytYnk8yrOI0m7IMg_X4LPb4lbDfF3X4XMYg,5963
474
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/boolq.py,sha256=rvwWqSRlFv-Lhc7QgVghe_iEpVtNx-2WbfysUF5TZwI,4020
476
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/blimp.py,sha256=FWLTaIqd4TuTBkLSLyu_ygHoPkbmvQhB-06mTJ6alDg,6656
477
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/blimp_nl.py,sha256=LlLs75FCg2sbxbULNiiCAGbU8GZrkZNXOXEiHGQ_EK4,5865
478
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/boolq.py,sha256=NA7TnsMmr887DKQoCRwgq84-veJ-VQdXPPfg09_IkdA,3966
475
479
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/boolq_seq2seq.py,sha256=-B7CD59m5m7WmL6ed3Ior9Zt57kl4jTOlbOkv9p_iFY,3998
476
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/c4.py,sha256=xjbcuPLg01i9SA5W21zusuhEUfnIaI3tMuCD4dlzi9M,5786
477
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cabbq.py,sha256=uFUWHDqytKtuwbs-iFB46CCvCsniNxq9j9j-Ae2nl-s,5938
480
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/c4.py,sha256=gangKka_3u7PoivPFCpa_iCSk2TfO86XcGD2ysUwiWs,5684
481
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cabbq.py,sha256=9qX7C6QfAmaepqc89FLkaXlCinNk7W1iiphSlfurZco,5840
478
482
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cabreu.py,sha256=_dS87XtZBWSvjTyzvBRjaK-dN0Fv6kDihux5PEj7neQ,5046
479
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/careqa.py,sha256=rKnAsp5IBez78ak85i2iQo1LliOMaweqB3GIyaBA9gg,6775
480
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/catalan_bench.py,sha256=EycMpZ-XGkRro3DO1xUZ3vFW0_jzDtMhcCvgALlBidU,6075
483
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/careqa.py,sha256=GDW61WgODlfcgcIhuP5BuCXDbQ-XHAtT4pDmq7yUGmA,6677
484
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/catalan_bench.py,sha256=M0Ms2i5mOJFw9ayF2qySwSTqrKDJwIVht2im6eGUWkg,5977
481
485
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/catalan_bench_gen.py,sha256=vYrbmDvEWetRTxFYa00e-EPmcG698rWSNTj6abBxUs4,3976
482
486
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/catalan_bench_mc.py,sha256=U27A3A9nSYqKmqSPWeVBU66eIDiw4Y7UjA0YttcELEc,4150
483
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/catalanqa.py,sha256=q6Hr2jQCkcVUuOjLyIzBYsUXSWQZL34352jXLcXMzvk,7256
484
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/catcola.py,sha256=1Qs6oyLrRYMDM_NHkCU80si6TddhiY6bsE5m2yWssAA,5509
485
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cb.py,sha256=CLBTHMpQ3RyY0gn-_rrhCgMWPNHna4bnHz6ldbkYkFU,4080
486
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ceval.py,sha256=8lcZUqajLrtuC_yDtHPmRgdwhH9A0YfoDQJrOAVQXA0,8640
487
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ceval_valid.py,sha256=euDGphqZuKuzsIcTw8cUw21NyMq8iZg8Ll208r0XiDI,6322
488
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/chain.py,sha256=pTCwNeM77UhiyrqN19d3NZLFImL3oZ9-fDs9Ael9hHM,4347
489
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/chartqa.py,sha256=BfKHnGYUSQuFeOg1NJK-2dmSfkV84FAE40Kj4UA7gE8,9313
490
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/claim.py,sha256=QenIkB_V5fYn_gACF0nNSWJu_wBumGPlQfhVzBs3dGk,6230
491
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/click.py,sha256=xvl2xuzFQMxa7gRrbh-pKfqAPF22LTqqZZQgQmv1asU,5938
492
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cmmlu.py,sha256=jvoB7X6gi0-tByNHlIMuIaLvQdLMW3WrsNATVuNWi8w,6661
493
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cnn.py,sha256=o3CkInmKzwZZDNFtNtRotFQaTnQrCsUjn19A1I8EWhE,5838
494
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cocoteros.py,sha256=M-MI6r1JkFFSTcQLg_wawf_QZEgODgNvbfroxr0T6PQ,6034
487
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/catalanqa.py,sha256=7GigYh1OrjB29b8E5UAbRZlWT2_IuS-ZZc8JyukRxgk,7158
488
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/catcola.py,sha256=Q2zzNtKybvSQAQlyAJusubSur64W1eFEVyqO5e0vFRs,5411
489
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cb.py,sha256=25a1mv3hRta1zna4r34nbUcaygCWLMXLm88XEa98Cis,4026
490
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ceval.py,sha256=Gx3Iw_TwdT_QK7i4iW0anY5IRjRMa5N09a1UlpKDXbM,8542
491
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ceval_valid.py,sha256=EDuO2CXJAeOE8CQylKY2MTKhH8NCwJFnP9WMZPyNskw,6224
492
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/chain.py,sha256=b0b7oclVtf9sqxlVODN76UFG5Skua1zxfyBs_bXy1_8,4249
493
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/chartqa.py,sha256=lh25ISyWo1LRUoIGrSMsOm90PJzdqVu2zoo6SYfhGAY,9215
494
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/claim.py,sha256=dKNl-GGM7Hb2KBMoJA3hng9Vnu3jGATmfgG4clDnn2U,6132
495
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/click.py,sha256=v_qtf7BlJYh3stuKA1964i1OWnF8DfEOvlKcdwi01rA,5840
496
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cmmlu.py,sha256=J1RwrSTizqoEoSBpR8gYQLbVVL8MBM48Jj7IOvIdcDE,6563
497
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cnn.py,sha256=rOBp8Lavy-H-x6C85C1CxZaInNDkxGGmSQT1yoHte4Q,5740
498
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cocoteros.py,sha256=zBoML4lLS2jSKtUKvgcdtmEEgemMSUaqDaAkUp52pEg,5936
495
499
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/code_x_glue.py,sha256=D14EUwniYnfUP7Z5mU5zRfkPAd1bCtYwEMd0tAavAE0,4511
496
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/coedit.py,sha256=jHan7PwlX3wyQjsW49qzAw-P7DlJPYjZQ-tEoNiU-l8,5875
500
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/coedit.py,sha256=iavlWyD1b2i5cRTJNiPVH7FzgadL8_BoHOVTHXZc3ms,5777
497
501
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cola.py,sha256=dRnZPR63cbrzVyiklwJX05cMB2fUTujZhNpQC6IVG6c,2996
498
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/commonsense.py,sha256=QLNyRHimX8aVUiNuT0K--I5-vYo-FNFTnvTfGjNNGJM,4296
499
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/commonsense_qa.py,sha256=FuiwbSPi5-iy48QHLmu01DCBnBziVZWVYCqJCO8NkG0,4565
500
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/copa.py,sha256=a7pe0QRisgu0ItvdAyKXZujks4Y-aofu-CnFEISKsNY,4141
501
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/copal_id.py,sha256=yyzr-c7prHDTVmcErfGdYiPnjVoKfrWm6rJUXCk87pc,6825
502
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/coqa.py,sha256=YR5hcHzOWZXiEjm5LjFlC0mZVAXfnk_NaZMKZgpoxE0,5790
502
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/commonsense.py,sha256=Z2TFnlZL--wWUWfRTaxHluiO3YgyPOB2T0BmvOkKryc,4198
503
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/commonsense_qa.py,sha256=dw_8F6yZxbr-Vf6v0yDNTqq4WdLOQQNEz2GALmuhuvc,4467
504
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/copa.py,sha256=xK9HQShXUwMxzLSaTRDfRgYEEDlKjTIK7HeFIXSBvo8,4089
505
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/copal_id.py,sha256=vK6mFf1mXehRsWqzXDDfr4MxLYRBv2pmSX5Qk6wgT5I,6727
506
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/coqa.py,sha256=wL2jO8QEO6F6XmVLKGbKKOcUAjisX0mo3M_899LL8T0,5711
503
507
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/coqcat.py,sha256=ebeI1P36JuauSiy2cuk8-BZahZfc8ylxeumXkh1me0c,4241
504
508
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/crows_pairs.py,sha256=IMDNwLFfzLxzFlF3TIdINstY4-XIrUfrMtUNmDn7g8w,6042
505
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/csatqa.py,sha256=2vs9t9MQt7W7qfWidHfXdX51jxr_P0VH_2D1hpXvupY,5947
506
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cycle.py,sha256=Yad__yL8WcQu1BddUd0uO0D1SQGrcOjFCCd_UZ6g3aM,4272
509
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/csatqa.py,sha256=4CX8ew2lcE60_CgEc-595xyc2YTS14H5cBjyLdBPq6I,5849
510
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cycle.py,sha256=plRdY59xzT0UYRLSw6v2_MZD31aHBNZIcxy9AwO5cn0,4174
507
511
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/cycle_letters.py,sha256=MUH1UEnt0ox2FCkDWrfatN1uoxNm43ut1SaToBYmJRg,3166
508
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/darija_bench.py,sha256=F-uJqfXHTKQuwuMidqf8GeFDRXzOIjamfKUKQanoKgE,9142
509
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/darijahellaswag.py,sha256=00z6_kfvpgTFuML1NGwdsNJ_GPwueuZ0VIS1ODcpdUM,6887
510
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/darijammlu.py,sha256=qtrgc7E8T1nzPAqQWbtpcpDM6Aa3yZSXI3VCEx5S0q8,5983
511
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/dbpedia.py,sha256=nQrrwGH9dX0HnI40czs5wa3_kLF7SX_clObmkTeqsls,6123
512
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/discrim_eval.py,sha256=T9fdZq-3lng9PhKl7C3B8MW0WVEXcz6tNT9GD-i3bhc,5999
513
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/doc.py,sha256=fZCZ4H4pob7hIAr0QIg2H_Pgz--TyQbVFYKuR8zxpoE,4264
514
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/drop.py,sha256=oQscohjTgBxuXA3oavOz_-VdC71YYUD3pNOpVv098cc,4361
512
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/darija_bench.py,sha256=I8DTdKhuNbNH2_BxCdACA4VO7hdn5zeEE7BW7KoOM8M,9044
513
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/darijahellaswag.py,sha256=zKa7oglGaoydRfFUs3ZdDye3H2l42aHU8dWOrEWs-gM,6679
514
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/darijammlu.py,sha256=yq2opVGKVbjSn-nN4scVUwBqrb3VYNYkwRyKblkLjYU,5885
515
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/dbpedia.py,sha256=6bW_HnrKD2zXP4FbHKyjeZqLYbPN1W9DfvYbjpTi0ug,6025
516
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/discrim_eval.py,sha256=kpv5zLpZDwisrGPh4DcDZ1iaVGiC0ahsJQhVp0C32f8,5901
517
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/doc.py,sha256=ZNq54kLCLbZWlqymCWxSTbrQQNSgYLgFCy95jXIxaqQ,4166
518
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/drop.py,sha256=Cm-vVcnTFP4FCxWspNCJIMslKTEOsZ2MeP3qgMWcJrQ,4307
515
519
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/egyhellaswag.py,sha256=44eoNICcU6-zYrmpUl4YGuJqSqBJ9KTew6Wzj-qfH-k,4577
516
520
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/egymmlu.py,sha256=Jtia9IzuQm_FU_3vk0o0BG7za_TNP_YvSiFk10J3IqA,6369
517
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/epec.py,sha256=8Cd2eZsUelZLNyJs8KDuo_d7oSHMtCLM7PV2OVp-mco,5713
518
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eq.py,sha256=oZmCA7Pn6bLDWtLXBT3KYpgDgxJOnxa0hZctOmLt2hQ,4260
519
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eq_bench.py,sha256=1IkAkQc_7xF5CEnhXQhcIxfC_XZahIK4n3RnWK5vZ_I,8045
520
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eq_bench_ca.py,sha256=4RgL2wIBPJEBb4ANJpD39qiFSeOkCeIGVXdrtMLIJQY,5988
521
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eq_bench_es.py,sha256=sPqsKjSUkszF5sShH5AekB822TX9UCCfZbMR215fqR8,5988
522
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/esbbq.py,sha256=3lvIcWmqVO7K6CZHa9cEv66XzhGTLwoIBHQNGl1_VxA,5938
521
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/epec.py,sha256=7_kFNMaMeBUNGv0sLo6zeRFm6NCAGmA9FuTbB7gYqtg,5615
522
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eq.py,sha256=oMj8G_mavAIPwneiAj45sEdZtGxCv-OnF_EYGd-J5H0,4162
523
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eq_bench.py,sha256=trjP3M1F8mCS-olXBCwYmOwnZCbRIh3e22Adt_paqtE,7947
524
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eq_bench_ca.py,sha256=FLWrFZ6YvKvc4YbiP1V2gTpxcHdwMO7qwyKgEDOA4oI,5890
525
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eq_bench_es.py,sha256=7frvhFFJZbbM1k51Pa7EF9T4RC2zhYP6aej10Y4CLzU,5890
526
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/esbbq.py,sha256=dhZU355vFqg9sxaum2oY5yE6xBluZyA-kYTprZw5lio,5840
523
527
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/escola.py,sha256=cDZkmuv0TXe4ShayNiLxMnaXP3B_9W7cxAOb-z3UO2I,3121
524
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ethics.py,sha256=hlwNlzBsNU__nYknMvBtmEv4Z71vbmQC1Soq19Gfoac,5571
528
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ethics.py,sha256=PR6tReNsGQTvsh0aiRXpVSW6X4VBdZemC3aE_5PldtM,5473
525
529
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ethos.py,sha256=TcDYU5eG0Ti4UumMVDt_RqjohJf9nESuxsXIhqyPCXU,3860
526
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eus.py,sha256=gjzA8xMvYRWjuGWoVltXlmJgi4KuXlQclbIY4w7shPw,4264
527
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eus_exams.py,sha256=p4NGqMd8S5vi97JbFGihhvp6D0PV7gihn0amVfG0cA0,8731
528
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eus_proficiency.py,sha256=iiLuH0hUypAMPcIqw91FsBOby4zqB9j8fJoj6EHLgj8,6340
529
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eus_reading.py,sha256=IbR_RZLs4qM5YRf5RUOOHj4N-UEYYuy6cn653iyYFx4,6292
530
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eus_trivia.py,sha256=ppuPnMO9_niyG_z3I4K-wojCI3vif1jHGQLbmDQiJCk,6312
531
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/evalita_llm.py,sha256=DW_u-xNGkF_84EtmpveX49Umt1RVeqEdy7qMq3i92CE,6596
530
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eus.py,sha256=tiSAOFOwVOt2vLcya-CHb2BuMvIVwUYrVnbP5LzFuFM,4166
531
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eus_exams.py,sha256=0vIbmOjebDq61eY55U53QtMmyL1c-kiMaRZElQHUSiQ,8633
532
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eus_proficiency.py,sha256=Ci966oYVevN0GvojhEtPwFVV41Cww0fIkJLzSAle6kI,6242
533
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eus_reading.py,sha256=TKpnuAYGBKtT8Zg-dXu-WCP1gdUBKBvWQn-tttYND8Q,6194
534
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/eus_trivia.py,sha256=rAviUzX3XBRX3DC1BzjRKJvkx-TSAj46KHs15fpi7pw,6214
535
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/evalita_llm.py,sha256=r0bOBF8hDE94f3dI55-wwFU_8mvHodsBE5h5WMzL9jk,6498
532
536
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/evalita_sp.py,sha256=rvmw3MJpPHln4p2FdA8b1n9lcd-x1wvBJcUALH5nHm8,3970
533
537
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/fda.py,sha256=WsQZGYDZUwb_Krhy9RjO6c1eSjIMdLif4osWB8eAq30,3944
534
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/financial.py,sha256=fkG64F2F8KBQDeG6BKd0ksWbl6RQnQkP9yeFNhzCRdI,4288
535
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/flan.py,sha256=MHFSgBXRp2vk9VW7IVYSy6wY64Ne1gzTa8kLyEddK0k,4577
538
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/financial.py,sha256=G1Q1R7OtQnNYPgHGhPKNDj82QprGs8urN9BBJP2JAGY,4190
539
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/flan.py,sha256=J5vWx4ToYRa_HHELcCLDDFhNnE3i96-VymOKMbOsdFk,4479
536
540
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/fld.py,sha256=GA7Jlu4XtE0nbS0lGCjx50lt8eJk5ZlT3bTtFyUX-v4,5216
537
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/french_bench.py,sha256=L-Di_tpAd4KjcKnAeuEZ2JNjjVBQI9fqhMNaUFkeQZM,7820
541
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/french_bench.py,sha256=T53Zno0Wl8YX-CUHa7ZEgxfxVMb4RZbzJ3yJF7ZzV_s,7722
538
542
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/french_bench_mc.py,sha256=H7BUHk857ioHCu9KhCjX6OTasagalE3_tk_Nw3Y1e2k,3616
539
543
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/french_bench_perplexity.py,sha256=ZWiytkj1TlCmKfouvpfJHz2dVmGMZ6ScDrOFXVcdPOA,3119
540
544
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/galcola.py,sha256=0ORvzlWMGi-9ybDKWSPG5pDe74gPTyigiHlMcB1fG6w,3960
541
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/galician_bench.py,sha256=KKGFFEmt0KGcFBtbZ1Usou62n6lAEuqIgoMrRn02HMI,6122
545
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/galician_bench.py,sha256=ctDWeenE6krvghsqse5tcgGUCzhhb6KioO1qW5P7Te4,6024
542
546
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/galician_bench_gen.py,sha256=3-GQBmOcYU_AH4kkVuudC-0i3z8zUR7gGMFH6SU8M0E,3956
543
547
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/galician_bench_mc.py,sha256=uuFVeFyNfEp-YUpNJshadahCvG8c5AAlr0ebxHAO7gc,4159
544
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/gaokao.py,sha256=KqAiVV-i_0m9cbEfMQFHaB4GKMTxsjeWK_cVLwyq3hM,4847
545
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/glianorex.py,sha256=EggmYhwm8vfY_6CZgpPB0SKEAxaxUgBOSJgYzP0F9Kw,4565
546
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/global_mmlu.py,sha256=GB-PacZ0kk96e7jxGQK0hHuu9u9cWI6CQWbeM3DpdZ4,6444
547
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/global_piqa.py,sha256=aWrPCIW_pOPlfLXtRbzQZZkEcVO9A7z4bnLwztyLJyg,5990
548
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/gaokao.py,sha256=o2ExOEuontJ6GXY6ZzTaZ5wMcGkYhr42gMNYDKDgruU,4793
549
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/glianorex.py,sha256=mR11nyPK30FI87QTY28KkchHX38FMOhfUjOS_ZYDgFI,4467
550
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/global_mmlu.py,sha256=FJD58ydy1HKlh1xV2lmu8_chrP89BvxafT2Z-2uunwQ,6346
551
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/global_piqa.py,sha256=SAHydNoxdWBTBptZ9GzaDVgMemeJxn4twxn03F1d2l0,5892
548
552
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/glue.py,sha256=zIk_KN6wOCs-1ZBNvWr2VAGaRK2-0rX2pmFBQYWyIas,3739
549
553
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/gpqa.py,sha256=OnRcBqQks1j0d2wz-Lwv1kt1HRdzo9v8djF50yWxO2Q,5641
550
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/gpt3.py,sha256=jLlJFYXCiO20kWI22OO64rxWjp1W9kjjD13ygx9DNE0,4354
551
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/groundcocoa.py,sha256=x8PbYJ_TvyeEKj33JD-CfyuR23jzCVNUuEX-zvj-8t8,7246
554
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/gpt3.py,sha256=gdCB1n7nQLgFbJ_8dDYmb_jRdMc74dx6iO1M251CqPw,4256
555
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/groundcocoa.py,sha256=wRRmgtI49KjYNeeAncqbphuRc3wz1PyIX-xRtk6PQnY,7148
552
556
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/gsm.py,sha256=JlretZyrsYcmJ0KyPYFWGLYFx8Cvt3-TvhgUOjahEmc,3830
553
557
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/gsm8k.py,sha256=zGkwC5Tevv7B8RBA-3D9EOykU_XLoJ7BNfiZrtX0vlE,4598
554
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/haerae.py,sha256=4AUa0l3HgFqj-Lw5UvmRr6gV2b29ZgLyngLK7MS3nTE,5947
555
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/headqa.py,sha256=6XAzuoqSDahyjX5BjE23zTUlv3KN5_PHlu0fjRIsrs8,3875
556
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/hellaswag.py,sha256=EFwXe4TAQsJtSq5U3I7yQK6_9GG_K2quG5PrQc-xJNo,4412
557
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/hendrycks_ethics.py,sha256=WA9FnU567x3h5tBfOFRJidDy9vTESCwq1-c-fhmSjvs,9149
558
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/hendrycks_math.py,sha256=TNDEF2Hu4VaXeQxpn0Pwt_0xKnpZNmjhhP8qxgehrKU,6657
559
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/histoires_morales.py,sha256=VbUKmBd9XAmvg2iJB9KG0UxXx_gEc9sFcwXkKDvuAcA,7236
558
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/haerae.py,sha256=60CPOeY7U3EfYtHO3q9jQzkxZaAy5laSTsjwFD4cQdQ,5849
559
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/headqa.py,sha256=Iq2e9XUwIQGMLDjI1xTSPEBwOMac2Qbm4XfiwGJh4_4,3821
560
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/hellaswag.py,sha256=B0LblxA_p9RvfrNU0xtnfmqMQD_d0gnap-HfiyQVcQQ,4358
561
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/hendrycks_ethics.py,sha256=igjBc-aZvR9_TtDWXb95_6DyHHRfaIROyAFon_p5UP8,8949
562
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/hendrycks_math.py,sha256=VO3w2xgWNmoP4F0wUnogeNdfpOKdWEBOniPy99Vo5A0,8752
563
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/histoires_morales.py,sha256=Dg99_Dm3_o5Aq-zD6cTnOyU73D1nenJIJLSwYfZuHVg,7138
560
564
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/hle.py,sha256=l0PNHf3taiF-UhE5VO7_GFXwtD9ReR9h1B9-eJ-ukyA,3791
561
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/hrm8k.py,sha256=DRbfEHav5FJwBgKL7bCv4uHe66MwwzwUtSitJe86-e4,7711
565
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/hrm8k.py,sha256=VQmK_k7LvfDCd8lIdDIorixI17hGwFp91AutrXC9A6g,7613
562
566
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/humaneval.py,sha256=FBNiL5EvqxbtnNurZDKAgiTb_P28_S0DVQN3QEKBGgY,4377
563
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/humaneval_infilling.py,sha256=DF0xtvaVnFD8BE1Y-NBgiVX22FxFRPHnxwL_ozxnNl0,6062
564
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/icelandic_winogrande.py,sha256=LAo6IvFinvL7-twDd5l1FK5Pa1Q99ZhgOETOV1dW9bI,6071
567
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/humaneval_infilling.py,sha256=byUM6QcTKPuRaYAWI7De9JX66hOVydYCod13XLaV2wc,5964
568
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/icelandic_winogrande.py,sha256=Jo30iNd2em2VqkMqw0wXOeBQMZwIa8aW9ObdkY-oCno,5973
565
569
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ifeval.py,sha256=J7AHYgnEKuePzvIR9pmMsCoaJNs6QIaSIt5MHLIHjUo,4225
566
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/inverse.py,sha256=Nq00uxvdezvBZhJKlL-T1GpAGxe-_r7PulU7kZ591vM,4280
567
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/inverse_scaling.py,sha256=gd29ZSQ33KZ_LLQP2QW7lqRK_Yq-fesT4b2wyReFewI,7680
570
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/inverse.py,sha256=n1PN49RGM4tfX8-UWPkaH--HhiIhued36iyPfKJM1e4,4182
571
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/inverse_scaling.py,sha256=K-8CHHaaLvw-EqBHc0Gn4Lnr8XDryhx4JisoV9RNzbA,7582
568
572
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/iwslt2017.py,sha256=noRk2FxmCjnk2RgGHWuvaoT3PqnjF9H-V1b8fogqKQ0,4444
569
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ja.py,sha256=3nOdcPHjGY5nvubXdGfuO5R9D6J261CQAqjvL2zBVgU,4260
570
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/japanese_leaderboard.py,sha256=XFLJEMJ36osS60OUzciVMXC4rtu1CO47dJrB18YZwc0,6150
573
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ja.py,sha256=9T8MhKPLo6xrUrxiAZZ_dqRY64kAymKu8fBj6sj69tg,4162
574
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/japanese_leaderboard.py,sha256=6IIqSgI-UUO-KbrP5rjvDBNNDLa9qCKI9tFJBmXLJHU,6052
571
575
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/japanese_leaderboard_gen.py,sha256=K3abgi4Mywn9BQMSQz_V4Hg8etbLHrBE-NpJa3LIivg,8857
572
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/japanese_leaderboard_mc.py,sha256=eIVjIolhbqineaukBxU2sf2EusFy6EjJ62gwKuu1zBQ,5037
576
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/japanese_leaderboard_mc.py,sha256=Ay9u6qetRNMElAD09upugrgVe-zEyHrGDxoNfHaMtCA,4993
573
577
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/jsonschema_bench.py,sha256=q-smnHB3IBpuSUebVPkcZGCqwZc6UEGD7p4MpvVNW4A,4687
574
578
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/kbl.py,sha256=ZTQ_2CF5Rj1xgU7bnZAVrDvv9F2sVORrCCmR9ZjOjEA,5179
575
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/kmmlu.py,sha256=yqeFyZ0-o8nCbrv8b8zcat3QadREiYAPV5fbrJNpcQw,6651
579
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/kmmlu.py,sha256=WaK0OCbwvTrLN9KNoaEPDNxvQ9rMHaazp2XNSQIuJWY,6553
576
580
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/kmmlu_cot.py,sha256=CvAVl1xMmvKOhpMd5INKXGjZs7wolf-hyLD-oBsH6a4,3291
577
581
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/kmmlu_mc.py,sha256=v2KK3DCmTkZyVmkw9gSTHElHV28AAp9Dji2vGizd_kE,4086
578
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/kobest.py,sha256=G8yiAzdC_o-1JC81nTYKYTbD9rZoz1l-GHmW8Hdk8oc,6625
579
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/kormedmcqa.py,sha256=LVjOwFoNedZVFqIPv0dubaonjUhdU5Hwh2uwYjG7CTM,5384
582
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/kobest.py,sha256=gyr3ovj-XiL-8mHLFXuKAqi7m8swhY5BmOmwkhj3aeo,6527
583
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/kormedmcqa.py,sha256=rCEJUWUHqLIILlFWwlGJsaWml2Fx2dl7bd0wdA_FIt8,4922
580
584
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/lambada.py,sha256=kFbqrlO0SSn8QBhVvFd4SvoQzr79efmHBrVOBgDM9VA,4892
581
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/lambada_cloze.py,sha256=9-IVLeyeiVQ8WhZnRL2pFJLfQoQGgdRYSxsYSLy2Hik,7210
582
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/lambada_multilingual.py,sha256=RnEBqMWvbCHP3kCfNlJ__ymWu6ogwdJyuPHaIBjZ698,7297
585
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/lambada_cloze.py,sha256=EWi9aPHaDCB-ENFw5qPtXuRouO5t5MFV1lEKmF_4PUg,7112
586
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/lambada_multilingual.py,sha256=pV_oP5ZPtulZOWbMvbJf3BCpy2_msFps81f9rObB1TE,7199
583
587
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/lambada_multilingual_stablelm.py,sha256=S8IBwsTM5yRzE2RRfjF4XakO9QxcIF_vmLbUceh1v4A,5157
584
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/law.py,sha256=96Hrld3W3srvQa8b1qWYEBbZs8p1EU2-Mn0Q1kbguVE,4264
585
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/leaderboard.py,sha256=nPRQZ9996-Jc1yrnElMyebbyZdyUA9ssnG2t7nhUqbM,8075
588
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/law.py,sha256=O9ppiN_atyGsui7s0_Z2yyW8Lw5gy-TaGsc70Qk9HE8,4166
589
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/leaderboard.py,sha256=qPopTbVEHUAlto9jdXKfOXnpB_kpdIZRB-zxAhxT0qA,7977
586
590
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/libra.py,sha256=le4wRHUc2Dpq0gTzS_lQKCZ8zBvlYMVFq0OMkzkdrck,6107
587
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/lingoly.py,sha256=nl_lf8-_pcNDSVbqkGaF1MPeiNYdnLa4H4qCm7rmBFk,7752
588
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/llama3.py,sha256=RiQLmKvG24XtH7ryrquAexP2Bqc1z16WV3gmsurrcAM,5947
589
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/lm_syneval.py,sha256=3N2Bt-aAaxugVmQGwt0MdQZ1qKzm8i_qWuqeQJ28zZ8,5981
591
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/lingoly.py,sha256=YIs_p-KVcRoumvs4QDSHV7dpHD5YBpTSMaMSsCZTUvE,7654
592
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/llama3.py,sha256=PCFYQ1atDy_8gWAG31zca-mtDtfqVyHm8nO1MRYbWyY,5849
593
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/lm_syneval.py,sha256=efyC-bmiBxJbcSNcNm2w23ZB7pt_TW5i6BKJbqDoaUw,5883
590
594
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/logieval.py,sha256=b8mWP5I0RUsmLSDE0UyYg9VC87C_qGQWn8Uq78G_dAE,2932
591
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/logiqa.py,sha256=tOAbI7oXs8MnsAT5VdPQuLzGfpwqzRfke_z0CDwLF20,3989
592
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/logiqa2.py,sha256=gRnnCQnDE9U3paCbsU0H1sjMC7Ubuil0E4bpBZLynC4,3900
593
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/longbench.py,sha256=5X73swbNvj7flqtFlNAqA-EGKX22QZM2mVeqtn2KW8Y,5974
594
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/longbenchv2.py,sha256=5f89b7cg4NHa_1vsIbFFz-osh3srP_4KMfbCzWsHc9E,5992
595
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mastermind.py,sha256=iyXStISnO5I9KBkgy9sh0nC9eAxmubZ4Iu0J8hIbga8,8019
595
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/logiqa.py,sha256=oBfvYiaEcWFnVRjqMbMgZtUbbOdQpRoQuRsZhuB2qz4,3935
596
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/logiqa2.py,sha256=EE8M804oNuftutfk2Ex-GHxQpU70_og24eDvkIYbXaI,3846
597
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/longbench.py,sha256=LSP40aDEXGFCK3PxkmMe28j4E3p49EoPFv4We9nLGQ8,5876
598
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/longbenchv2.py,sha256=B9zYsogZNQKWAXrjA03tQLtijXX42fnmzwohrsFu9nw,5894
599
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mastermind.py,sha256=ZF1dsVeI5HCtgf57XXkIW2KCMo7aiaWjhcUt1Yxo3V0,7911
596
600
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mathqa.py,sha256=Pd-E5D6pRWjN7J8rMVy3IbnU4DuEd-20tAyHrMkxiYc,4779
597
601
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mbpp.py,sha256=qkCmDbZcE5gOYMMeJrvoxjtZxq9wa7bHiEGpi1Sp1EE,6035
598
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mc-taco.py,sha256=rSvVod1RH3CWYzZi1hH4MJOnYvnFYFwhOSbwQhckSJk,3955
599
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/med_concepts_qa.py,sha256=QbCTKrOXGmd6oP7Pt-yRa8idy5ZcIROuXLsjsSeouRM,8785
600
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/meddialog.py,sha256=ld97Bu1mJneVcQ1RlcEyJvnRwxLCJUfYReHKSlMwv9Q,7082
601
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/medical.py,sha256=KzHsg8KVHTawJS7C5nLIYGw-LsFgq2ihr6NZ-bgFYDo,4280
602
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mc-taco.py,sha256=JLSBGNYt7oCT1-_-IE8y4MiVTxWgoEw7PAC5TXZr8v8,3916
603
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/med_concepts_qa.py,sha256=74Br8u-kBnBTHh0LhoqVXHJ6vMBJGPtXfim33A-a9SI,8677
604
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/meddialog.py,sha256=ilYrl8A_YWPGZS3A3-leh-h57tXr4gdqZBXNg-Rc_f0,6984
605
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/medical.py,sha256=EOrfPnejfA257_Fn9CuXWgf9-0_8M2AunW6DmjLLrOs,4182
602
606
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mediqa_qa2019.py,sha256=W34U00Bk3jwolmds8-P2X-3fSULK4Eo2k0hACs5Qf_8,4796
603
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/medmcqa.py,sha256=_o3_lWxSjmF0vh_JqnjJDKRp2nkN_ukUGDziKwFebdo,6711
604
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/medqa.py,sha256=J1ixlsjUUccQNCM-ShGaSTntBkBUX1lOObkV5a6Hnwo,4026
607
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/medmcqa.py,sha256=Sz2cFTbHTzKWa9xxWrVnw0XkAm0xlL7NKmpZOXiXBAI,6613
608
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/medqa.py,sha256=-aOWbDE6OzeXbjBiAWAnaR5v-gmsAsPRd3j32k78xhw,3972
605
609
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/medtext.py,sha256=OeKPSgBPfz7XjboeW3KGVmSq19xjAFp6dYz0tASwujA,4143
606
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mela.py,sha256=FXPKqXefxdMTftO3ob6Xx6Yetm9ZQOgENpHFsEOg5B4,3828
610
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mela.py,sha256=uhfXLPsl8bGp2PpepmlGOBZPU0PV4xBMBZDNVVvrUMk,3774
607
611
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/meqsum.py,sha256=C1DZEn4peQm9GVfMs4fE0yKkMzwCXReZ88rsFzDSigk,4259
608
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/metabench.py,sha256=ooghyQ-N5KCtlCFtGP91V_hKQEkSD_fWjESu_GSxwFA,6041
612
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/metabench.py,sha256=_fBlmz9305jGNkEYNCbA_r_03d-WZMq2z45Mgkh2_uM,5943
609
613
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mgsm.py,sha256=K11m-ohBZ5Pxflp9ZaL8rXIeHbKmpHj2rX-uba72lcA,5060
610
614
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mimic_repsum.py,sha256=VjHLEhKwuHkHo8s8gnLg_5qR5XE3wdnizdzxQVGOG64,5351
611
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/minerva_math.py,sha256=UovLnmz4vU1mU9-HGBiH_WzPnXBPkJAhrg55mo2T3Xs,7213
615
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/minerva_math.py,sha256=pUXHtXk8qkOFKOKMuFN8fzqPjdko5i9a585OwWjwO24,7115
612
616
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mlqa.py,sha256=VSx6TrBBLl125o4tF_m6o8e4huRInd8-dKs4lMkTdEA,5659
613
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mmlu.py,sha256=J-iEuRvPmY3_ORP-d3l95TV2TWvo4Zqn81qW0zzzhOQ,5475
617
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mmlu.py,sha256=wtnv8mUk50QhOKOgjq7T-OTv8UUMFhVt2V4bGpanY1E,5377
614
618
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mmlu_cot.py,sha256=5wtrR3gLHlhf6qy_yA7rQZOl6zxumrMbd59LqMSw7EM,3284
615
619
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mmlu_mc.py,sha256=MI_L-X1QK9OglTbxNYRhkL-nBRumVXGPk_nLRZpt700,4079
616
620
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mmlu_pro.py,sha256=A1xG34WxJmIkrRsGtFX_Bfr8rKFIjFl3kLlJnYplO9w,5353
617
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mmlusr.py,sha256=gJehtqOZkZtcPytpshr6j0d_rmBiHiZpOzWawySw_Fs,7988
621
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mmlusr.py,sha256=XKr1C6Q_Z1e0wjcQnxp5q5uhoiy5xILeWGwiFFuDBX8,7847
618
622
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mmmu.py,sha256=3Hh2-BdiHZpz528vPniN9f0Ac5dhuoo0J6_wRg2grN8,5606
619
623
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mnli.py,sha256=tZvZzsyp2Gxk18KEECx8uX2JLvB4VFKyJgTI0S6CGLQ,4241
620
624
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/model_written_evals.py,sha256=y-NGZVAl2AA0mN2kaFwZoNV6TouR6L-3kAWatld5q74,4591
621
625
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/moral_stories.py,sha256=cUK-aNA25X-hJsv415dz_J8I35ChcinYMYAPMbP4y08,5957
622
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mrpc.py,sha256=XMaAyhwhM7UCnf_rr474x7C42DXHpKftWdDlpYH7p_0,3820
626
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mrpc.py,sha256=Xr0putTvWkF0K-e1yJ_7TZflcaJG0ziyPnMr7uqJK7E,3772
623
627
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mts_dialog.py,sha256=x54oU2Wr1cg_7X1NYO-gBSV9iEbI9pRykEwR9Zlsg1o,4308
624
628
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mts_dialog_perplexity.py,sha256=o3ZsLcemHjPb2qCJlv7i7kFKLFkKKbRIIshjwmPTdcI,3622
625
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/multiblimp.py,sha256=EbtO0PwvTIycRpTKNpayw5OB1wsj3MrghxIpHwdpd7s,6260
629
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/multiblimp.py,sha256=WRQn49f2dV_aF0H5qoeoQiM5cFk9KSnKRwwYL1rmGpY,6056
626
630
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/multilingual.py,sha256=9VBd8kU21XX5P733a1qf8JkID1sdo6m7myZJJcq1kJc,3722
627
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/multirc.py,sha256=QNkhgWHvfWnKTgQ9WqIYgpo8TwMo4C78MlaqRlzuRts,3929
628
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mutual.py,sha256=OaGST2AovGTnuiKj7dLL3Uoo1I0u-7UZoW3TmkVJwOI,3860
629
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/non.py,sha256=3Ruriwz_Tb4y1OroMaO8KWZmV9tEcB8MgQdU2N5uMU8,4264
630
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/noreval.py,sha256=kZE4fRbg097KGSzyCF_bsOCDg4GooFAOTItWWpNUn0w,6973
631
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/noreval_exact.py,sha256=lKJF-H1lV8-_X4imnfWGs2IAYd_77yAbGvfmGAb5Egk,6140
631
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/multirc.py,sha256=euSnkmCq0i8kZULtd36kSr3Eb-cfX2RGU37oI46rQLw,3890
632
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/mutual.py,sha256=-74jVzJ5_Fk5WImwjSHjIBSXfxG4NZl9WT8saZh2hE8,3799
633
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/non.py,sha256=pGsL8VhdqMCRqhYz4Z0fczdebv8vNxeYcSnuVuxMB1w,4166
634
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/noreval.py,sha256=B1PTCgZd6b_LTTf4pAorQuxubvEjMQtuZzmr383M3Ic,6875
635
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/noreval_exact.py,sha256=r1k-0QnTMhvLHffjSd1KEa-vDGQ7YbeG3ax9u23t9Ms,6042
632
636
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/noreval_gen.py,sha256=KROJgvdvAtd_Otrnss_EzPxF6upZiBmRRAYzvxJzVsE,11344
633
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/noreval_gen_exact.py,sha256=zi3V5pNcbIev9mPwOwhyXSATm9oOcvn3xMVG_MPW1P4,6472
634
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/noreval_mc.py,sha256=VDxUFDOhsnCvabkl4D6Fda4qfTDJyYun0jYcUj94DvM,9603
635
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/noreval_mc_log_likelihoods.py,sha256=mMVdFEJJ5vM63kC429DZzPbT5Bupp322ZuUYjpjHBco,9392
637
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/noreval_gen_exact.py,sha256=VHXF9nZypig9k6ARHqLj7EVjM1dNQI6vpCuQudWiRyk,6374
638
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/noreval_mc.py,sha256=UUmYAQ3h9Rp2HTZN0Ahel9bx_mr-QWRFn1j1p_rMEEA,9347
639
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/noreval_mc_log_likelihoods.py,sha256=vGVHM16g4CzZkFe5flLJxTIPbaNTQe-GcLZSaERYU6Y,9136
636
640
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/noticia.py,sha256=fBvmsoMiGB-N5FZ3t39ZnTW44cKPIrXYpKZ-iBrWLjs,4896
637
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/nq_open.py,sha256=KZAuaRKcm6-askmTkmJzbl97CbzeB3rxOieIxGLVjBs,4632
641
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/nq_open.py,sha256=pBHE0l28uUO0a9vuo_w5oZM6OAmhqp5bVLUTYcaY4MI,4578
638
642
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/okapi.py,sha256=WiABcd4en7SjqkTsKq9ACWB_-tPULj6myzrjqNCevKk,864
639
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/okapi_arc_multilingual.py,sha256=JKBcUO1ItfE0IHxEbGvLlhZaEpwHJ6bqNmbUKzsd8rE,6835
640
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/okapi_hellaswag_multilingual.py,sha256=vvqHq24sVxjmQablTlR4zI_9JVp6msl8a8gSWwzha-Y,7424
641
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/okapi_mmlu_multilingual.py,sha256=UI6gwnQL8LwZmIxzatwVjF39CPs-UMdu3uzq1WFQmms,6618
642
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/okapi_truthfulqa_multilingual.py,sha256=m254nxIhoqRYlVXK4R5w2ynFqDdidPRKEYgvlfq_W9E,9505
643
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/olaph.py,sha256=NySRIqiWaiqJcGxoIDGo5Lefxkliz2fcrMSrPd8Hhfs,7279
643
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/okapi_arc_multilingual.py,sha256=-j6fmTQ-atZwLdX9A9h7ATYbum59fs8Z86PelUpe2xs,6737
644
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/okapi_hellaswag_multilingual.py,sha256=XZlWDfBQKfGT6i9ZvilsTdDaB7vusrkXon_Lzp3mAqc,7326
645
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/okapi_mmlu_multilingual.py,sha256=hBQc0zzirsYrYSPqQS75ogDry11ESli1tAYN2OPrwWQ,6520
646
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/okapi_truthfulqa_multilingual.py,sha256=cIluL4k-oM8GGfJjbAolExdiBfmmiG3MUjPucmlH1j4,9298
647
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/olaph.py,sha256=7Z5MpTlpyQNcr0Hfo8znYYEy-Nx9bP-1iRQaN8Lt9mM,7181
644
648
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/olaph_perplexity.py,sha256=3aff3dr8ZHWw4_FRVqweVQNL16VmHokKlmsoTyBSuVc,3594
645
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/openbookqa.py,sha256=HDfNLMKVT-2mKOXzEUO7kBeMe_VxSHzhfM6VoiqyKJw,4108
646
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/option.py,sha256=xT2bpvNzLwHpv_VmiE25Pvh4VSCYvh5LGUMGFRnIchs,4276
649
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/openbookqa.py,sha256=qEYzfmDme-kv6GoA2y6bh8mPVZLi1rSx5KKX9Pc_nj4,4054
650
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/option.py,sha256=UvciLPmH-M2o-4W8WsyiMnpkIpCyLRxy0tSWPMC_Uz8,4178
647
651
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/paloma.py,sha256=Y7vKEU051LHAtmzb7lAO7LgNm4orDmWLb0D1AVBxrHA,7904
648
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/parafraseja.py,sha256=fk65rmGff9JHwUVuF9PwdRlS40z_bTFS4YuUKsqOMEc,4366
649
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/parafrases.py,sha256=vu_ttWB_ZQolk81am8IuCiweK4qywNp3mZQJqvspSFE,4364
650
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/paws.py,sha256=RxPeMbRdaKfcIjKqeA-NyOyGoaZTL6k142cdrr9LDBI,4268
651
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/paws_x.py,sha256=LNpE0cZ_Ox4EBLgLWhjjoLdQ2UnmTNRL-q-8DRe_NnY,6009
652
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/pawsx.py,sha256=YE34VO0dYrXLgSMbTpgif-HSM7uXoWJDQlud7H0nPc4,4055
653
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/persona.py,sha256=t_QLyi4v0wbXk13B2Fd1pJCDlVrAXT5RCOtuaZsIIWM,11055
654
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/phrases.py,sha256=E7I5eEpN6-Qc4bUDj1GzC8hO3cqatFrZ1oRDl0gF7GQ,6130
652
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/parafraseja.py,sha256=Aw2j-xLLlyGKTJv1v-SWAK0XFZuEmPqVF7sDVXQy0OU,4268
653
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/parafrases.py,sha256=aPWJeY0t5oEanOrtDx4-KhBGNZmGyUZW0lf9nHL10jQ,4266
654
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/paws.py,sha256=k6hOYMIAHJ3AlAcCv3LHnwO--2P8KEwfY2bD3EBeJJM,4170
655
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/paws_x.py,sha256=bA5RBuW_7QlRQ0ocnZKWxUjIr2s3vFgGBBuj0OoJq2Y,5911
656
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/pawsx.py,sha256=UB1mAqWRmSsjG3qQ8taVztSDWeOTUOTHW4QSJfMaBKc,4016
657
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/persona.py,sha256=TWPFa2XT1AQm3bZBGfQybN1_rnUAmKoxZnuSZY_H7PA,10957
658
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/phrases.py,sha256=36-y7OkqgCDaRNZoI1eDI47gMGdYQvIVn-PR9ztaJL0,6032
655
659
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/phrases_ca_va.py,sha256=F3p14GK8kpSsknnl88JYFY0ezuZmgWit8iel7QBtIc0,2986
656
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/pile.py,sha256=cfr9GrsLPX4XXAORV3PkXpC4_2hKb7u24PswVLAMnYc,6432
660
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/pile.py,sha256=I8ckihfe3j1GS6N6L8sL_Yc5eyL1ZiKay7invbCcM4U,6334
657
661
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/pile_10k.py,sha256=K5diNndk9obpncxS2h2pJ3WeFpceZTlCO7d1QXeKCYA,4845
658
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/piqa.py,sha256=iXN0CxFzLDvLt8pWyuV9MlOUXIiIIkN7obg79eNPtnE,3884
662
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/piqa.py,sha256=oJiOSW9B6mcwB4vBbuULMHBrgG5FPa1Rz4ZSZ0KWyMg,3838
659
663
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/polemo2.py,sha256=8WYXV84sPbNLxwfnPLSVTDKY_suFlJPLp9ffbs9wnZ8,5179
660
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/portuguese_bench.py,sha256=Jg3foz6stbTDqr9Fi2B1YUd2ktH_PRvfi3-VfCazs98,6105
664
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/portuguese_bench.py,sha256=A9x_WfzoVetzN7lctFF3sHAne0CgxcaouggZvk-Ug2c,6007
661
665
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/portuguese_bench_gen.py,sha256=hGLtogrXR_yib9riVQZAkqpwCsMbhb7F8K-AR0uKFgU,4191
662
666
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/portuguese_bench_mc.py,sha256=u02jIQXUOuwbjfCrDtd1rIEvPq4e9CPvvsa7De8DSYk,4133
663
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/prompt.py,sha256=DBUw7iKa56bniQjGOYi81q0fusND2A6fVufeFpFbjkE,4276
664
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/prost.py,sha256=-Ub_F46PhmsBCCP2r4t_o4LpPnlyBAWljcBfGzoCn-M,4060
665
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/pubmedqa.py,sha256=HKZ07Ysvyhj3nqGS0LgtYD0IB4vRzjHMC5420FRpzB4,3971
666
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/qa4mre.py,sha256=5HUAPE93IX-0wbkNBpg7WZI27u2Ki2snF8BgbpZ8cDs,4154
667
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/qasper.py,sha256=IqngDM_SscCLYfoh9IlPGuJqGdL7MByIvFMBUF7ZPHQ,3961
668
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/qasper_bool.py,sha256=b_aQQMVgr8GcMuTmRemXYbvcJ7bTCCHe-lF9g-s8jPo,3839
669
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/qnli.py,sha256=t1OUysTdpLEawtGPNdPLu88P2PRcum0BuC6webyC1M0,3797
670
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/qnlieu.py,sha256=Ku4MpOlfBOsnP5utbA78UecsYpoL67C_jqv0EOEtr_Q,4276
671
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/qqp.py,sha256=qw1E7XxC5rOfDNWDysrvKxZkJ6k4iaw0OVEg1AmMCnk,3796
667
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/prompt.py,sha256=0u24XuKiwfGonvzaZe30-TBZixmDI2gDJAPcUyuLvrk,4178
668
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/prost.py,sha256=FckuTbXhBhxwJ1MZfL7I6CiIfLnIExyhz1xbXDBYbnc,4006
669
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/pubmedqa.py,sha256=eJWfNKxW-yjPQ1GvN3waO5WjKUQoOal-vEX9RHrRETA,3923
670
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/qa4mre.py,sha256=EiuU5mCn_HFLbYG-qbMt43h6jrJz-xmAHLl5Lgtpzvc,4100
671
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/qasper.py,sha256=lvKfMu49ZSuLgSfmoURMIL9itWxK0n8mRrK7gtPEWIo,3922
672
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/qasper_bool.py,sha256=ZeDBo-byUQuHDV0WYgPZ9I641rBp7xZgN2GjmA3jzNA,3800
673
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/qnli.py,sha256=Bgl3kpP6kWQGFDvB-pZl0uESF0c3N1i6ufHdpYXx1PI,3749
674
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/qnlieu.py,sha256=JX9aJXGzZ0p2Uv1VSnbU5krHoNBNLt0aGuOE_dZdb4g,4178
675
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/qqp.py,sha256=lk2tFRew4E4f2O-DxnfUKOib8BiMnDXqEQy_pu1rxqM,3748
672
676
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/quac.py,sha256=M7HVhIic8RB3LvV0LCsDdMqEZCPBK7sWb_fkH-Y8HGY,3797
673
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/race.py,sha256=zjZ5liczuQiSSwYFJWtpcgaxEj3Fo8kEf6gR6hMcSOQ,4141
674
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/random.py,sha256=krT116DEDmYHd9D6ym2RxswrUD9FWOtaYuEVS75Wj0w,4276
677
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/race.py,sha256=QlFL2jsBV7XwoVo4FiuqUqBYE-6TYboCSQT1xschwqA,4087
678
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/random.py,sha256=mFaACpCJMJhNCa2zj7WZWYRkgPxQAvQ8apaJBFJEnlU,4178
675
679
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/realtoxicityprompts.py,sha256=1r13H8HW4N4J9eSDmFCAmdoKGhShjxjZ1PCJzhSLTcg,4575
676
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/record.py,sha256=Rq2W6-7Wq3HhvPkY_psn5rEQJ5wXWwvxzZx8gevQupM,4261
677
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/reversed.py,sha256=wyqkUVJKm1Z4a08iRtyWqc8oZu-b7-uRCuPlnNqGSBk,4353
678
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/rte.py,sha256=zv5RTqERoz0VrxHMD4BldP5K7o9CZNq6gqqqQRjarLc,3785
679
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ruler.py,sha256=7dCa8wZ1QaC2OjQ8XSyZYj4KxVnbQgI48oL4a6VdZHk,6829
680
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/sciq.py,sha256=qq9Bns5Luby1FgoHQ8E_zpWXXiLji8iVtcO112426Ic,3907
681
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/score.py,sha256=gVsOLD3_elWMEJpcnuBC834HSW7Qrdgo8tzCpIcLfDg,6787
682
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/scrolls.py,sha256=EBWbrmRFYp0dh2Uu2nEFzivc6g7jV1nNPJ0jhaGhpBg,6131
683
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/scrolls_mc.py,sha256=k0TBC17INJMgPlzw3oP6B89kN8Z6WQjvz-GI_FU_NQ0,6042
684
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/self.py,sha256=wo40Mk6a3mmR3vnVEhMMyMh_Z-lvQtJHhsTFpxc2vtk,4343
685
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/sglue.py,sha256=Cv4UwWLmmHjgghMRKjVDCX_NLahZlwEYQEP5ctv9fzI,5368
686
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/sglue_rte.py,sha256=r2YKxLs5c1d_SAwuvKkpgY8Xme-_BrMSPRi51g6w6uM,4180
680
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/record.py,sha256=IElbX-S3xDVEjnDOtMAgq4-PtO7tUbXnLfrotzMehUo,4207
681
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/reversed.py,sha256=uy40cjl6s-E89Jg40GmdK1Rw1NEg2X0_9LHTTG5_ROM,4255
682
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/rte.py,sha256=sj4eviWxZo5Cqyo4CG21-DBa3ZeKR8RnE-mihLlv1bg,3733
683
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/ruler.py,sha256=t0UTzPFmnLYt82X1TeWzDdngDJV5HVWNu-TFbEdXg14,6731
684
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/sciq.py,sha256=brMvk3IkbUbqnVzoToqjIjcLWye8ipEG7JuFufTYtJI,3853
685
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/score.py,sha256=D34PY7yyMzyYZrDnTHRfsWSzb_lm9zeSaSqvPSzXN_o,6689
686
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/scrolls.py,sha256=4HgI2jViwUAIhIXs44kEP7mD6-C0Ju7BIS7YNfXB2QU,6033
687
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/scrolls_mc.py,sha256=lMNuCvmRq5MZa9Xd_EfsVite0qDhoN6nD3RfYpPRBTs,5944
688
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/self.py,sha256=NrPDLq_VjFL-ipnKts2MQGr4JDnhf8gk5it768xTPlQ,4245
689
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/sglue.py,sha256=FVFIMSMHZyVD1PFIxRq1PpHIgGj1BQQIjN2QK_MkvNQ,5270
690
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/sglue_rte.py,sha256=eHTrAnSwEeBTvb66XJQYYAxgIoyjRzwr39WY3a1NPx8,4191
687
691
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/simple_cooccurrence_bias.py,sha256=pipWwa2PNQVTfJeqU1xi_lPbv1XVUdgedKXsku2Lrpo,4490
688
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/siqa.py,sha256=9sB7reIkFAvekmnTdqS5jML3dTMrePkTY4YyaN97Qvk,8229
689
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/social_iqa.py,sha256=r5BT7cy-1-N-qhwBJGERDXtok9UqTGyZ_v69FiTOJTA,4055
690
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/spanish_bench.py,sha256=KD_S6j-6s0VTU4JoZpJWAePQe8jX-sf61GwHbGY8q_A,6075
692
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/siqa.py,sha256=2GqKb6ccsjWn5wzPrSmH9xmIfurxX3veocy2vmEGxFo,8028
693
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/social_iqa.py,sha256=O3GSl56G6wUZ1iYRQETcY0gvfQ5l2ifReAVLvtyduG8,4001
694
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/spanish_bench.py,sha256=MaaLf_FbtiF0DxjTDyY_LsihaSMx77paPAZ91eLwCKc,5977
691
695
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/spanish_bench_gen.py,sha256=N878fSFoeO3334gnht8zg5Gfb_sy1lbsZcsj0D5ENck,3919
692
696
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/spanish_bench_mc.py,sha256=izYHAzjLmdFVfmUMB21ZZZhiUtDIMUdyKo9yNfPm404,4128
693
697
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/squad2.py,sha256=vMQezYU6gN0aD5oiLItMCjRAxxXa1_GBYc6VBCSeJMs,4602
694
698
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/squad_completion.py,sha256=qrkVhWAw-gKGfwNCqglqFP1TCU1_LPJQmIr6qd2i2iQ,4410
695
699
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/sst2.py,sha256=eqpA1VNI08nMdxF4ltNq333RJLICvENxSS7svQY57Z4,3851
696
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/storycloze.py,sha256=NVsKcygSoSvFPK8sR7OESS7Xo77fTtDP7zWp-tCRA9I,9610
697
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/summarization.py,sha256=fN0DBEbhmoddKzTLiQfeZepxC3BXQJ8g4BpVZoHEAYQ,4304
698
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/super.py,sha256=sU1s68McNYfsNOOTy5awKECOfUSQTVloaBPDlrq1HVQ,4272
699
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/super_glue.py,sha256=7VKmF7tz6b8NKsqyHUM0HtciO47AHntQlBfI2tLfRfE,6049
700
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/storycloze.py,sha256=jphtZcg9VIYZ3krgknrNtPQIPwEM5nV0bdEag0tTJyY,9406
701
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/summarization.py,sha256=TWOeNTmE1lfa2ks3Ook9uphU3Z6jBFa82vS0kTyiBnc,4206
702
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/super.py,sha256=rRp5VEcz7_ltUWTJOUwL_gPiNZzX-OH-qYP0r8ZUviM,4174
703
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/super_glue.py,sha256=hdKqmtrqaCiSfLqCTIPp9xiRc65HTDJVLKK4rpHVdr8,5951
700
704
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/superglue.py,sha256=NosWWhOPRF45WkNCgJFJXI4x132tFwRX102DcdrXf0E,3827
701
705
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/supergpqa.py,sha256=4hIJpmgdVxZf2xsFOkbvgIkfyFAXF0vbnwlCNyoYNJU,3827
702
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/swag.py,sha256=2Dx3mNKV62o2hNSqzdLXc3Oa0A6ddxEBCAZxMAy6WYY,3946
703
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/swde.py,sha256=uE8falrDdx6_mAIAHRAYmqieQPATtiyVBu0n8Gu3pTM,6978
704
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/sycophancy.py,sha256=IqJcLdsXNIbjIZKbELv5r6_XWKFUXTtNnbzpwHzla1Q,4375
705
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/t0.py,sha256=F3V7vUCS4-U1_OS_yQPxtz2E3yGTckJHUKjW4cyHKCc,4321
706
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/teca.py,sha256=DSeFGcbi3AzZ6svgHeUZmkaGOIgh1GwfOuWRsOR7R5U,4331
707
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinyarc.py,sha256=dxHaJWhITTunfEOZ4ko9rWMzSo905lc4aMpIMBWXr14,4346
708
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinybenchmarks.py,sha256=N1OIJjW4FCC4nk97PUB1Snf_SQrkSOBI5S7KGIqnOXk,6092
709
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinygsm8k.py,sha256=EoVAdXegybLiH3alutR5Mwu6v_M7QtshRGUtsBrhJ4k,4351
710
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinyhellaswag.py,sha256=IBscRNFLP6Lw4GdzEDUnF99hdjRB6gs0PqQiC0yCluk,4376
711
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinymmlu.py,sha256=8kHFHNkXAEaOC_vCsaikK2VPPWBIMzV7Q3Abq8aOZgQ,4351
712
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinytruthfulqa.py,sha256=UyYnf5BrrVG6RmqqQO00FpZbs048bzqi3tXVRjHY1bU,4413
713
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinywinogrande.py,sha256=aHPaAnzUNJf4hP2CHR1tgx4cmQpDMIW5TyZGVY3HIPE,4381
714
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tmmluplus.py,sha256=Om7XEfXXiVUt6fBt9pxgzB-rq58a8HwEQIrt1uR3iS8,8206
706
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/swag.py,sha256=0ZB8PvzKVhZiZaq_HS133xj3-5MqucudKZQysq6IVLo,3892
707
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/swde.py,sha256=0m0WMPiUMzjDfc1wuLFtpBzDU4l3twvPLl8lg3R1dH4,6880
708
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/sycophancy.py,sha256=ldHakmbs_ChOCzgfAF1EtOngoy9edmXDdpmhuZlkUFY,4277
709
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/t0.py,sha256=2DKa1BmS6iZLLcN3Z-IoJsYYXiJlrEnAR-5Q6HN5h9Q,4223
710
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/teca.py,sha256=VDg_po-TKD8gwF8naWmOea6ec8_OO72f0BhRtSebYKc,4233
711
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinyarc.py,sha256=W0usZ9riBei7km4WV8_eXTnfL2IamGls8ZegisqBlmk,4248
712
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinybenchmarks.py,sha256=8CK5WIiWLQte7KW-afuRKrcK-MIZggf1Z00BJsmt8dY,5994
713
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinygsm8k.py,sha256=GLP3EObZ_9j_-Sf62ucgwHNkmsRHDG9ycR3CP5_mMlQ,4253
714
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinyhellaswag.py,sha256=3dGc6qhDISKjNACTU1bXVJToQHlsh756AD_kwtVGxlg,4278
715
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinymmlu.py,sha256=4JyBUFl3O8Gdlh6Wija9JV2lUwEhcEQr_62rJWtCAK8,4253
716
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinytruthfulqa.py,sha256=El-fPGj0szWZlRK0V9SjCySm9llvQQObVXzg0Ri_-gU,4315
717
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tinywinogrande.py,sha256=3_-EFzulS2GxI2E4RbmGd6MYHdhQoy16jlLtvlX_lFM,4283
718
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/tmmluplus.py,sha256=ItfpZdnxKoiONGUHiBZMmMKRK4lS9FoK4s6UVZjWLCo,8108
715
719
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/toxigen.py,sha256=wyr9CDq4LupITsRR327vGhIfcgjXdtDgcZBPS0EexDc,3270
716
720
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/translation.py,sha256=rHNpXKyLCl5qUtTw0nMjg0jW4sFOK9SOKkawc0UBH3g,5167
717
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/triviaqa.py,sha256=-R3_JDrAIDk68ejFptAeESvsx_OxTNbTNEmke0mFnLw,4414
718
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/truthfulqa.py,sha256=QE3LhA1lDiP-DnYBceul4p5mDc5uUmml7dXXu4aKXdQ,4480
719
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/truthfulqa_mc1.py,sha256=orxcmYBJkZD0SmH_9tHULNXTxe8VVFq0tQE2cF-a3fA,4102
720
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/truthfulqa_mc2.py,sha256=bgMmg2nHCk_t-Z_KsgcuQ8jc9JiRhiakNyzbm3z0bYw,5096
721
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/triviaqa.py,sha256=6Hc1IkZXfBtDAN7qrcFtsLcssnPnKdCXZRa6c1HHHWM,4360
722
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/truthfulqa.py,sha256=nO4LRVeTJXKkjXvDRZJ3WK4onJSYRNKx-kmYqXSIdvo,4382
723
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/truthfulqa_mc1.py,sha256=70Tuawb5FTrhWOKeAezoWthVf81A5FrATeiIX-_AhGM,4004
724
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/truthfulqa_mc2.py,sha256=PW686J9_AAdPyLBlsYeMQ6XROczPb1IcEjeTP_f6g6Q,4998
721
725
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/truthfulqa_multi.py,sha256=oCWC6hY3RI9Nj04Rij71fDZCT-L17kXirkGE9fyob-I,4945
722
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/turblimp_core.py,sha256=L1T6pDwdy8s14F3O-cSxd5XE6rGU3qVpEkV1JiuWjJU,6008
723
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/turkishmmlu.py,sha256=C6_VNlabW0JUH55Ls8Q3gu9QvrlRsxvZlLtZCSxxDC8,6338
726
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/turblimp_core.py,sha256=Cg3edq6WfVd6444Ap6YUEmKNiU4mnaqFv_HzZ3g4QEQ,5910
727
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/turkishmmlu.py,sha256=WN8weWS1TNyNy3qw9coM_vknHYSsCnb40yVEq1CvjOI,6240
724
728
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/turkishmmlu_cot.py,sha256=nD7B0tA1rMd80eazkrnGjXS1VemkEp-pvqMweyGMphg,3871
725
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/turkishmmlu_mc.py,sha256=DSkDuI1WXm7KAUUOkE86O9_NMcQHwIpnHQN6iUQSC9Q,3790
729
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/turkishmmlu_mc.py,sha256=sSMEhTucmkxhtrqSD85suP4ZjIOJ14nk9uTNDvmux0k,3702
726
730
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/twenty_newsgroups.py,sha256=_mAy170d7m1SlPtbNNkToyr5zFasvyEUlWgJyuNZK38,4563
727
731
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/unitxt.py,sha256=OhhhQJfwBTSjWipi3nPIelpjrSpvI3vCZ1FmYPH5WGc,4592
728
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/unscramble.py,sha256=B7B9-pXNQsLlwqrh_CSBSHyqdGfhUgTzrEkj6FzIPf8,6048
729
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/vaxx.py,sha256=kJ_ZoftHm-qkwXlk3nIexFzdJtuY1UywNp4haNg9RSg,3581
730
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/webqs.py,sha256=qX3-nvQzZ4VV72bS2AUuBvgHnZEpwZGxz71_gUSAqOY,4438
731
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wic.py,sha256=vGI2GxJb6vfGJssxMNKMngprAI0P4OMDdoE2DJdpWG8,4108
732
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/unscramble.py,sha256=3LyrHvUVZwlas6vVriIh28vJUYVKzM32_UnlBRst7G4,5950
733
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/vaxx.py,sha256=TfckjIy-_sV9_tLjPxmgpdPa0kO3esG0h7Yf_o7ms88,3527
734
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/webqs.py,sha256=GHzoWid4Epw7-lL4oygxfoT_oaBK_kiNAoX86MGzKVw,4384
735
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wic.py,sha256=HEE7qyKyEMoJlXUKD88iWpqrsCA6bnVsbSxiwkgDXtw,4041
732
736
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wikitext.py,sha256=r2izxF9w-LsQg3erIk3v8vGQttc7nxpqi0KHkYe_cPc,5000
733
737
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/winogender.py,sha256=1CoZe0Btrt06gu6TCfcPnobbewPgY2EOdijKISq0bgE,4766
734
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/winogrande.py,sha256=rynWcZsS17eb3mpGU7PifpRPzsJbhFeOKL5-0g3dC00,4114
735
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wmdp.py,sha256=1CTyt1f2ABCsXLtKgobwWB0_89q_6Sb7b_06Km7EqEA,6030
738
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/winogrande.py,sha256=deTEUw0oT86cl_8xq6rqEbPKGjZnOazDTcYZAlE4QDw,4062
739
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wmdp.py,sha256=OWnnlt6Ng5RayIxvUsB0HstVFti_-pROBxzaMNfnQkI,5932
736
740
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wmt14.py,sha256=_zHPAk_V9VyigtVcpTLpbLgFoNrHeASloddh9lQLUGU,4043
737
741
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wmt16.py,sha256=pX8F91Um3kUbsKgxJZDdRBfFzlmCr5WUx-OFBYGF5CM,4446
738
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wnli.py,sha256=kR3M4BF6AyVOYmY3nXdwHQHYLwXkZJrulrB5FAJ5FQ4,3855
739
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wsc.py,sha256=FTqa0mnmFfpuSVUOuLVosvBJOMpBszPiiGUvUybKLxM,3974
740
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wsc273.py,sha256=Q-xGA95r9l3qLV-PKBdXaLrkD6AIQmI5KU21ZrKJ9EM,7074
741
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/xcopa.py,sha256=9c6CfB4wgdkwroe-RzOfudR3Zhcp7wF4HnEpvI7yaZs,7463
742
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/xlsum.py,sha256=qlh0yIpvpcVBgS9-VXTcmVGyZf4wpiAXv3s6ELdZFr0,5371
743
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/xnli.py,sha256=kryrRr9VJYaioi-c5cMg-kbF4l29Q-fqW0Ri2HkHIC8,4181
744
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/xquad.py,sha256=GE1qp2Pc_FDTYFCt-rSRWi6u_b6hXQvvH1n4l26eZ-g,7745
745
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/xstorycloze.py,sha256=11GknEiqca4grspitZFZRYllt2IFLqL9mWW3ZIyp-Bc,4530
746
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/xwinograd.py,sha256=FO-LW5iMLNO3SMk2EKtWVyTWrpNSOsqum7qp0MqNHgI,4108
742
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wnli.py,sha256=ZjvaIwrqGk1CiP9Q7j0-9LabEnov47vpR53tcbGhslg,3803
743
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wsc.py,sha256=aY_bzRuWwDu2sUb_YTXafZvO8c1_9E5E5m4aP9dCuLg,3932
744
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/wsc273.py,sha256=yhslwubim0colqpT6t84EOrp844RfDeRrXFKDT-kuqw,6976
745
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/xcopa.py,sha256=-xGIiysi_c_mht3MEZwX3dPabfi7kYnTXNM7x9afrPw,7365
746
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/xlsum.py,sha256=fViGmOE9bi0o5T6K5yXwdHfqZ7B-FWhuecFdpuNVCps,5273
747
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/xnli.py,sha256=4q2n2eWy00QAU238ruA057f8pfwOYp4dTfUKcZmAEcQ,4127
748
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/xquad.py,sha256=VI448Ro_eDF-rzPdmQlCuoCkQIFWs-Hoe6d_sC30NOw,7637
749
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/xstorycloze.py,sha256=gr39S_zUyO0llrS_y43j-S8CqgIj86saTBpsZwIyiiU,4416
750
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/xwinograd.py,sha256=2lsXdPSVTzcuhGWK8Gx8bXmRh0xTqWcoLRKA3KrDBW4,4054
747
751
  wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/yahoo.py,sha256=oI2VKhoGTIVPrJGv4L7-o2H-8CauckVDql1wRMs_WWw,3949
748
- wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/zhoblimp.py,sha256=JGAocj_5n2zJJ2H3Z9aqB7P43hRUoykf0XJNykqD9aY,6032
752
+ wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/zhoblimp.py,sha256=U13gF5-nYth4L9gIIlbpLlbvA10dXYbUwjPKcaUXUO4,5934
749
753
  wisent/core/data_loaders/__init__.py,sha256=VAcOq6wRL3oihqeJz6Pk23u1enwEkM8w2lShv0V8Sik,8667
750
754
  wisent/core/data_loaders/rotator.py,sha256=hlahoJz_cvl7pxsFwVDcF1fc9Uj690rb7aWpqdfq9HU,3794
751
755
  wisent/core/data_loaders/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -753,7 +757,7 @@ wisent/core/data_loaders/core/atoms.py,sha256=26vJ1B35_8EoQrFOEdNvXHcJRGxSfqsYpF
753
757
  wisent/core/data_loaders/loaders/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
754
758
  wisent/core/data_loaders/loaders/custom.py,sha256=Xe1sOHH3_dRjBnQg9vwMM-XA8ROn65dUr9TeT-nuNtQ,4144
755
759
  wisent/core/data_loaders/loaders/huggingface_loader.py,sha256=AvMG2DpEAOMfzz339WlDJ3qeqLjvcqY7P4P7lVnjKTE,5147
756
- wisent/core/data_loaders/loaders/lm_loader.py,sha256=GZyawEOXVMaJSPuD8nkGhNsmcWJ1XS3WzfXBAEQk8d0,62200
760
+ wisent/core/data_loaders/loaders/lm_loader.py,sha256=_uPsvYXGApETcvn9hooT3rDdgrvE8gfXw3bDyewKL4Y,63150
757
761
  wisent/core/data_loaders/loaders/lm_loader_special_cases.py,sha256=-Ku7dOZIBgJjbNAbP5ihXvRn5KCStzSmk8_nLsshmzs,18885
758
762
  wisent/core/data_loaders/loaders/task_interface_loader.py,sha256=QGDfqMvdTFfogF9P533b-Gw52uHe0GnsS64AfnVVlek,11771
759
763
  wisent/core/errors/__init__.py,sha256=DX-T4zW2PwNvTcyXyt1T_-oZ_xXl4WhzqaOZKiDaWFo,5535
@@ -828,9 +832,9 @@ wisent/core/evaluators/personalization/difference.py,sha256=8D9azBdqciYo4IJt2lwr
828
832
  wisent/core/modalities/__init__.py,sha256=jtowcmTMQ_kOYai7FKwXY0vqSYQ8vews84VSMH73XFM,20138
829
833
  wisent/core/models/__init__.py,sha256=G9JfaVU_CXCG_wnz5D5gjDJ-_G7zqHFmjaixDpbCixM,456
830
834
  wisent/core/models/inference_config.py,sha256=0n8647avNb4w0o0UsbWnU_eh6cMU23ppdJ-FDYx8mAY,3704
831
- wisent/core/models/wisent_model.py,sha256=ZOiT9WfkeZneqfiu_aS8XSIg18As204GR1j28R2D514,36526
835
+ wisent/core/models/wisent_model.py,sha256=sweOiwBMsPaA1AZ3zGPn6ao-oI2kbbtTs-T5DMkWDd8,36566
832
836
  wisent/core/models/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
833
- wisent/core/models/core/atoms.py,sha256=7QMaJHl6axH6Iy0woZFsgEsBga1IaMhVq7q7JoASRic,16107
837
+ wisent/core/models/core/atoms.py,sha256=XsasyE6Ki7jIwmsn5n1r9sOHPhOpKGdMfvD_MRSDSx4,16170
834
838
  wisent/core/opti/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
835
839
  wisent/core/opti/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
836
840
  wisent/core/opti/core/atoms.py,sha256=7D6TxhEcBKxCplajICCfpZQ0g2rsKB6iWoRB0Mwr7yg,5438
@@ -842,7 +846,7 @@ wisent/core/optuna/__init__.py,sha256=efbr9QeNaeWMle9pppWBJXs5ObiPti9QZBQGBvuli-
842
846
  wisent/core/optuna/classifier/__init__.py,sha256=vv2wCAbw8ccZxq3vxrQt9plUdbr7VJj-t2rRnh6jBR8,819
843
847
  wisent/core/optuna/classifier/activation_generator.py,sha256=9zV4qEGVovBP_3HYQq0UXArZ_xPSIdjHEiGHzBveikY,14432
844
848
  wisent/core/optuna/classifier/classifier_cache.py,sha256=Q8PNBJ2rpHWvAQRVPR_ax0zvnrbRKN0yUjOhCwXf9jw,17439
845
- wisent/core/optuna/classifier/optuna_classifier_optimizer.py,sha256=T1K6CSWreQeX4X-AeAT0jaf5hBDp1bP92LN_tDzm500,25490
849
+ wisent/core/optuna/classifier/optuna_classifier_optimizer.py,sha256=VmqEGNjM-omYEQmsneQs7uhZGsdOPfExSCTd4XLogH4,25511
846
850
  wisent/core/optuna/steering/__init__.py,sha256=C1O9Y2AMFsFvLq6fmTW_wJ5vnRXKVGftykRw3NaiHwM,508
847
851
  wisent/core/optuna/steering/bigcode_evaluator_wrapper.py,sha256=ZVp159jbcdUAf8A2q5AwINak0n6YqBUuq0iDLvXObyQ,7157
848
852
  wisent/core/optuna/steering/data_utils.py,sha256=PUQ8Gqu8oPfoGGFi_lTfaPXhTZ5MRQO2Gf73vRa7NDU,12431
@@ -850,7 +854,7 @@ wisent/core/optuna/steering/metrics.py,sha256=6UbW9AMrRpT1pgvDVK9rb1-ieqqOq1I1Kd
850
854
  wisent/core/optuna/steering/steering_optimization.py,sha256=V5iu5uttoT1Uq6Lj5jiUfnJ9-vIXVyAVD6UzXHdvXWA,45016
851
855
  wisent/core/parser_arguments/__init__.py,sha256=MpBsnsPcZQ8cqAn1tN4OYfb3oMQraNW-23g3Z5gSqeg,299
852
856
  wisent/core/parser_arguments/agent_parser.py,sha256=PG3V-F9xKE53guiiQ2GudGKZnjvtoAbdyMI8Qhm_sZw,4971
853
- wisent/core/parser_arguments/check_linearity_parser.py,sha256=HnYdSYOOP9MvVyRjO2ZvTIlKU-y_Hw3yXL5F_ehjZXs,2013
857
+ wisent/core/parser_arguments/check_linearity_parser.py,sha256=82a6WJCUdPyM95rb4NrtUMhd-XQ2lNYq3NByxm4k4Yw,2392
854
858
  wisent/core/parser_arguments/cluster_benchmarks_parser.py,sha256=QvEjOIRZswffCNMcEsIgCt1a7V-CzaHbFVWTaDF2-7U,889
855
859
  wisent/core/parser_arguments/configure_model_parser.py,sha256=LmsS0mYF7eP-zpYB7SdEbtFZLVtPtcbkEi9CMkYqMVo,350
856
860
  wisent/core/parser_arguments/create_steering_vector_parser.py,sha256=GvyHt0fqwVk5dDmAsilOttsybZ3jj5Z6dQWzE8EUj-Y,1735
@@ -863,12 +867,13 @@ wisent/core/parser_arguments/full_optimize_parser.py,sha256=BTd2Wc_qUkrdwvaPeoXk
863
867
  wisent/core/parser_arguments/generate_pairs_from_task_parser.py,sha256=wEZLIXocqgLduufNMYTNm4YcQwerq53zp5YseCqiCbY,924
864
868
  wisent/core/parser_arguments/generate_pairs_parser.py,sha256=nDTQ63bvjl-GcSrMdtcR_Pdh1dzbqtRnGaSCFhPO-4Q,1918
865
869
  wisent/core/parser_arguments/generate_responses_parser.py,sha256=LnjIx7gOdysEMGwUIHFHlaEOE306qr4rXto9gzBSjkg,1345
866
- wisent/core/parser_arguments/generate_vector_from_synthetic_parser.py,sha256=Dje2Xuq-Xmd3HVYzDCluifqJfW9jTZ5ccleUyS8ZpPk,4209
867
- wisent/core/parser_arguments/generate_vector_from_task_parser.py,sha256=88gnq8_fmQAmn3zJR1q77rIBFiOBz21nfyJQmd_H_P8,4873
870
+ wisent/core/parser_arguments/generate_vector_from_synthetic_parser.py,sha256=MvwLJYm-23pAwAocJqrjWS12RlAMW70ATCEWa2VfJ2w,4225
871
+ wisent/core/parser_arguments/generate_vector_from_task_parser.py,sha256=95jOVRH22aFcngJJYrLvpj-SFAPZaTSlAeh277NuWbg,4889
868
872
  wisent/core/parser_arguments/generate_vector_parser.py,sha256=bf2PHAdvNZetXaSe3sKarvK6McIODPdmRSNRQN_rHto,3622
873
+ wisent/core/parser_arguments/geometry_search_parser.py,sha256=-ytDrseoIRJnjMq8pvh6jE9fxn5IXDCVmZB_5wlub8Q,1739
869
874
  wisent/core/parser_arguments/get_activations_parser.py,sha256=pCBET7CsFz-8WFE3mb677o4W-vVa09wz-3YDMx-8ukA,2445
870
875
  wisent/core/parser_arguments/inference_config_parser.py,sha256=-TAcjy8yTzDKWFc42JCmVtGq1MowQI-IauRNFNj_se0,1853
871
- wisent/core/parser_arguments/main_parser.py,sha256=BKQOWO-6l3USgiJrldH_tZ9x-io2grjK-UaYw615hQI,12345
876
+ wisent/core/parser_arguments/main_parser.py,sha256=6RPGpcf5eXqMpSK4adVVU_vJit3pICZMoAzgjpcb29c,12789
872
877
  wisent/core/parser_arguments/model_config_parser.py,sha256=RlcSyQkj0uDlmJJE8sMq_CjvJ1VdCcjf8_BqspsSHd0,3210
873
878
  wisent/core/parser_arguments/modify_weights_parser.py,sha256=077Gz6GLk1aR2k6rboU4tcVBgxHaTB01sSRx7ZqlzPU,9336
874
879
  wisent/core/parser_arguments/monitor_parser.py,sha256=qo3vyyVmdZBkhGuhlHqbEzTiIKcdFIvG3IIuc-DIRQE,1234
@@ -882,7 +887,7 @@ wisent/core/parser_arguments/optimize_steering_parser.py,sha256=fRG64w-eb0ezUp9T
882
887
  wisent/core/parser_arguments/optimize_weights_parser.py,sha256=7SEOhgBs9GWGLyPzNfZHQztt4o24HHUPgCf7sbW1kVo,13904
883
888
  wisent/core/parser_arguments/synthetic_parser.py,sha256=99fGg_k9vQl80COuVZw5LOYLRQ0sDlZ4ozwsx0T--Qs,5135
884
889
  wisent/core/parser_arguments/tasks_parser.py,sha256=MNSijAG6zzinuwy9lIOwJxrL0JavM7MKx443Nwe7uEs,22440
885
- wisent/core/parser_arguments/train_unified_goodness_parser.py,sha256=5BICCLDk6UrUjDXiwsOIYJYx9ubPwwPhC70gAP45Rq0,4962
890
+ wisent/core/parser_arguments/train_unified_goodness_parser.py,sha256=DQF46sBOAvYCdzzdQlGvKhKgsyZPiR8sdZuCmonuWV0,4977
886
891
  wisent/core/parser_arguments/utils.py,sha256=XTtssIt05HlAyO6kMPZvu4l_UIZSMDaV-gl50NO7qOw,3543
887
892
  wisent/core/prompts/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
888
893
  wisent/core/prompts/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -901,7 +906,7 @@ wisent/core/steering_methods/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeR
901
906
  wisent/core/steering_methods/core/atoms.py,sha256=kDN-lvaQnMi9UTCdz-4rZ8x5baqrsVGVYGO-P2EpVUY,5377
902
907
  wisent/core/steering_methods/methods/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
903
908
  wisent/core/steering_methods/methods/caa.py,sha256=EsodLtBV66YR7ROngU7-9fJFhzWcmzo7HFjZCNE28Ko,1717
904
- wisent/core/steering_methods/methods/hyperplane.py,sha256=ST44J2fspz1sFkda-w5dAJSE4SHRG4ujfcnA0PKnbGo,2919
909
+ wisent/core/steering_methods/methods/hyperplane.py,sha256=NeLbFoxlDrnJ_8PTQ6J3BGXtNSTRbaLpwXtpifNgCP8,2976
905
910
  wisent/core/steering_methods/methods/prism.py,sha256=Cb18ip7tl57hz7KW6br4EsLWxYP5__bMYppsyBEPzIY,23995
906
911
  wisent/core/steering_methods/methods/pulse.py,sha256=aowQtMONcBkNaO99gT7qVClifKyyEdWPQqKMv-PwjSE,25801
907
912
  wisent/core/steering_methods/methods/titan.py,sha256=OVj02CrOlO_YVKxiOK6kBZUiViLZL3ot2709J91tFf8,41521
@@ -922,7 +927,7 @@ wisent/core/synthetic/db_instructions/mini_dp.py,sha256=XElWJGkIG9DR5q4wcY11kJDJ
922
927
  wisent/core/synthetic/db_instructions/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
923
928
  wisent/core/synthetic/db_instructions/core/atoms.py,sha256=jBazuD37hyBMIsbFNDjpwlGnxECAOuh4YT0fmpH8C6o,748
924
929
  wisent/core/synthetic/generators/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
925
- wisent/core/synthetic/generators/nonsense_generator.py,sha256=M92jIiXmXBows_FdF-cyo2JgvKQIWRmKEN4RacsgKdk,5790
930
+ wisent/core/synthetic/generators/nonsense_generator.py,sha256=JqgJKUU6prgT62R5OzNy-8DIAlgkWE9jRp5uExrqnTo,6156
926
931
  wisent/core/synthetic/generators/pairs_generator.py,sha256=VgxeXwTe_Ctvms8b5oBZL-H5l2ck9rExfQf0FHxeIVE,12927
927
932
  wisent/core/synthetic/generators/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
928
933
  wisent/core/synthetic/generators/core/atoms.py,sha256=9wL0v38BCqn3y9LtoRkQsK_X3egjdYcPmFXH0mgFSWg,2290
@@ -946,22 +951,39 @@ wisent/core/tracking/__init__.py,sha256=rda2gaV_tFP6CChzx8TBQUrh_PObqvX35-BHJCxi
946
951
  wisent/core/tracking/latency.py,sha256=uKibYRodOurQoCh5MqfJwo_k5Ho4H5dFQ0p_nddHGOg,21909
947
952
  wisent/core/tracking/memory.py,sha256=Fwg92c_M997uToNLIZvFlHbp3F5TTrUZsA_X25J-QfM,13201
948
953
  wisent/core/trainers/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
949
- wisent/core/trainers/steering_trainer.py,sha256=yJ44B5l0jVHeWKQ2yajSGjKvBLSEIUfdUo5SM8Yt_ow,15105
954
+ wisent/core/trainers/steering_trainer.py,sha256=hTtEP--odC_SfXaVQ8O0i66f6isDKyN5koeaZ5WMxUU,15091
950
955
  wisent/core/trainers/core/__init__.py,sha256=zkv6f6Auk4dbJ5PTQP96gP2TUM55W3tc1flHLWBvI4M,1246
951
956
  wisent/core/trainers/core/atoms.py,sha256=ycWk0G-7EIAEOnQL-o5_V5B8KTQ7CQUilGF4ibjighM,1536
952
957
  wisent/core/utils/__init__.py,sha256=ee4tMv5W_qxtTwRS-n8ZCl69CgBUqPGjugrkomOhzhE,1419
953
958
  wisent/core/utils/base_rotator.py,sha256=5seiluDYXeOARs7Z5RIHlfW__9nUbxir6KGCs0wUHwI,10018
954
959
  wisent/core/utils/dataset_splits.py,sha256=1G8TTiNFbjLcf17LjVh9xxlAqC1oyCMnOTQckHxPN6E,5937
955
- wisent/core/utils/device.py,sha256=zqen3KLYV8j5qUyryCTv5lxTW4Ygn8dzQSt_pgUW_qs,9419
960
+ wisent/core/utils/device.py,sha256=Ohe-WbVUDHJpOBn8NtxkFJ2lDgklokDk3Oi_HlNuZf8,9283
961
+ wisent/core/utils/layer_combinations.py,sha256=3zyQqxiSJM8vpJOkU5Y2AXewUjESwmeLeYU3v6IA4l0,2462
956
962
  wisent/core/weight_modification/__init__.py,sha256=UrQh_lVfQIAddYQDOkmVWY3BP4X6uuvltHY6KHPDu0g,4329
957
963
  wisent/core/weight_modification/additive.py,sha256=USpIfiz2K486qisbs8nubhyfaGgMGzT4CJnGjAB17sU,11375
958
964
  wisent/core/weight_modification/directional.py,sha256=GLcbn7IRQp2uqgl1t5uRA0IE9QukfyiZGNxt-YVdK4s,49253
959
965
  wisent/core/weight_modification/export.py,sha256=zikHU9dGDswa61ZPd_xo61kAdNt46O1VOQm3HOSYfNY,10266
960
966
  wisent/core/weight_modification/multi_direction.py,sha256=tpnEG54GHIzWmytt3ajUmwTbl2_ClpZS7AebJHHrkfA,12858
961
967
  wisent/core/weight_modification/utils.py,sha256=h6QM5FrpvOu29_6YZNvs5bwBuJBvVgtBo18RQ_MW3_o,6907
968
+ wisent/examples/__init__.py,sha256=gaDcawBaga7CMOkxnu9cTGP2-Xq0bN1IkWZiDEPPCoM,18
962
969
  wisent/examples/contrastive_pairs/humanization_human_vs_ai.json,sha256=7D6GpW-B99nXnhLj2q7WYZp5DWu8AMFzk3uyq7ambV0,247507
970
+ wisent/examples/scripts/__init__.py,sha256=HmFOUP_4AjzolMRK2FXhLzz_u29L-GxpppYUjW7M-Kw,30
963
971
  wisent/examples/scripts/benchmark_tags.json,sha256=JjZGitGdgbJPXxk6PXGDsEaQBvAKlwJOAcET4rucRWA,29279
972
+ wisent/examples/scripts/count_all_benchmarks.py,sha256=WjgyFCV07QgOQfxV0_04GcQYi7eRaEfb3CipPYjAcHM,3913
973
+ wisent/examples/scripts/discover_directions.py,sha256=3KJjksYda9UB_GrFV3pNmgTfTi3u_Gxz0zpgZkncyD8,17532
974
+ wisent/examples/scripts/extract_benchmark_info.py,sha256=3JME-s-6UUBuQRBOlM5TgVyMu9pkLbF4jdm_QB9Dfuk,2243
975
+ wisent/examples/scripts/generate_paper_data.py,sha256=zKHG54r-gYtqvdhJJONXjtBsGtVeMpOEu2RbFOG-74I,14467
976
+ wisent/examples/scripts/intervention_validation.py,sha256=mrxM2E0DZgtxMuXi_LPJ0NFyGDSU0ZMwC1qk4I65Pv4,21962
964
977
  wisent/examples/scripts/lm_eval_readme.json,sha256=X5bYjQrAUw031rjjFlrTlxxnllWGBSlqywF5kjh4mOw,89869
978
+ wisent/examples/scripts/search_all_short_names.py,sha256=brLUqAKTnmo4KA3v3vnou0BWOXUK2M0qpofPSXC0du0,1191
979
+ wisent/examples/scripts/test_all_benchmarks.py,sha256=oKwCgbSPQcQwRrjRjWdWn6Cz-JRBrRiWu7svwnWJc-g,4207
980
+ wisent/examples/scripts/test_all_benchmarks_new.py,sha256=4E1K_IuzC7Ut7dVYdd9lHZqa6dMMlsHPcbTz7JVhDCo,730
981
+ wisent/examples/scripts/test_contrastive_pairs_all_supported.py,sha256=FIQ6P6hXfLBArBp0OhliDaeqlbvRJafJiO9dMjBIkmw,8011
982
+ wisent/examples/scripts/test_nonsense_baseline.py,sha256=spdWAUb7sViAGpHOJyu3HThABcsJ_eolDxyhV5fkdBY,10254
983
+ wisent/examples/scripts/test_one_benchmark.py,sha256=_2iy1GB2CYXooZCuSo7vJ0sKBtjLZofsWSIO0lkEtu4,13624
984
+ wisent/examples/scripts/test_one_coding_benchmark.py,sha256=AXoD_E6kaNnTnI4Czi9ZYhd6aLrmjZP1PKDhKIUXb_A,10427
985
+ wisent/examples/scripts/threshold_analysis.py,sha256=Ugp5Whn7L_7C8S8JrskfdtOXIRz31Lt1232Uva77hgM,14906
986
+ wisent/examples/scripts/visualization_gallery.py,sha256=DP025fuVXEJ8-CPOB0O88IGQfek__TuHFo56Y-zfNv4,20243
965
987
  wisent/examples/scripts/1/test_basqueglue_evaluation.json,sha256=4dmo2Mu7zZ_mFhrn1EWn5grPkOhNI4XjKUYmhCoICl0,1898
966
988
  wisent/examples/scripts/1/test_basqueglue_pairs.json,sha256=yCLswgp9q3SDzhso-pB5jbIlTHzqS_GEBtwCQgVwlXo,650
967
989
  wisent/examples/scripts/1/test_bec2016eu_evaluation.json,sha256=rVRPrb-pot0bCu1p69ks2ltQiW_grUtRvxZonYupUY4,1897
@@ -1023,9 +1045,17 @@ wisent/examples/scripts/2/test_babilong_pairs.json,sha256=OOWUdyqtMXUv2_WWIqTTEY
1023
1045
  wisent/examples/scripts/2/test_bangla_mmlu_evaluation.json,sha256=OtR2JvARwR99QphQDprHiyv7wjLPK9583gR_rCM9RQU,2073
1024
1046
  wisent/examples/scripts/2/test_bangla_mmlu_pairs.json,sha256=GR9eXDwYdp0CBOyOZ9XF_yHv07KlDcpv8BpPF5wovXw,1340
1025
1047
  wisent/examples/scripts/2/test_basque-glue_pairs.json,sha256=yCLswgp9q3SDzhso-pB5jbIlTHzqS_GEBtwCQgVwlXo,650
1048
+ wisent/examples/scripts/results/test_AraDiCE_ArabicMMLU_lev_evaluation.json,sha256=fcQJbCj5AxAT0mHqSXH6M8stbp54KpWyKJw92qaPq8E,14062
1049
+ wisent/examples/scripts/results/test_AraDiCE_ArabicMMLU_lev_pairs.json,sha256=inaWlBB2zpzRyO7-YB1VYH0AGsduByBJ3_kwN4yTac4,6024
1050
+ wisent/examples/scripts/results/test_aexams_IslamicStudies_evaluation.json,sha256=zXXxnIR6Qb7y-Qh0WE-mrS7Oq2cZ082dUlEx6KG86bs,21734
1051
+ wisent/examples/scripts/results/test_aexams_IslamicStudies_pairs.json,sha256=XtRDKooN-Ht9Vqf8RRQPVdOyGmh1ZAODoxmQhLHNto8,8668
1052
+ wisent/examples/scripts/results/test_afrimgsm_pairs.json,sha256=qNbvIDdJ7Dss-T5bmSCxuUZ-e7UpjPStq7MbTjE-Sqc,6400
1053
+ wisent/examples/scripts/results/test_afrimmlu_evaluation.json,sha256=skwDyoaU4zTpk_Qg-HFAcSKR3R2OEERRi20xFKy0CmM,12575
1054
+ wisent/examples/scripts/results/test_afrimmlu_pairs.json,sha256=uOQ6MK9iXNt4c4h5S3-RI1KdBCV38T2FMPBWhqb4FEc,4851
1026
1055
  wisent/parameters/__init__.py,sha256=dwyW8F6LSMPM-H4zQX6gn7x5qPs78oNAMYRiRPJ8UCk,36
1027
1056
  wisent/parameters/lm_eval/all_lm_eval_task_families.json,sha256=MAFcazf3UoDplgaeWdkcKpAOzFRHyc5yLoXLqOwyFvc,2555
1028
- wisent/parameters/lm_eval/broken_in_lm_eval.json,sha256=C8vWqqJBRSUnLwCnuxwt_hfpaykfkxXN8NRJUCUixzs,120
1057
+ wisent/parameters/lm_eval/broken_in_lm_eval.json,sha256=v1sON6hTeKxORXVNrW_WJsIyFWhdIkOWwEXX2V5u2XE,3223
1058
+ wisent/parameters/lm_eval/category_directions.json,sha256=AfgWHCrVEvsjqvFeWDocY4hh3jlYVqkfhrrztWmKWTI,3555
1029
1059
  wisent/parameters/lm_eval/evaluations_not_lm_eval_tasks.json,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
1030
1060
  wisent/parameters/lm_eval/evaluator_check.json,sha256=XrfAOzyw3SYSIfqOo2Q_U-nXpM13ck4m-aRtL6Za9cg,131833
1031
1061
  wisent/parameters/lm_eval/final_verification.json,sha256=clscpavGxNXbX2Xn-idHKXGO0_fhMyDalV5bYZtbYiw,740459
@@ -1036,7 +1066,11 @@ wisent/parameters/lm_eval/no_readmes.json,sha256=T1PNoYwrqgwDVLtfmj7L5e0Sq02OEbq
1036
1066
  wisent/parameters/lm_eval/not_lm_eval_tasks.json,sha256=zcdUnstl_OA52DO5Sf8ZAPkNcu9B-bfuyTXxPm8MEr8,1964
1037
1067
  wisent/parameters/lm_eval/read_tasks.json,sha256=MLOSbTPLXhMcPVw5QJosIIlUiRTFeQmbFZ16wFcMMKk,3401
1038
1068
  wisent/parameters/lm_eval/readme_files.json,sha256=1VtyYzREk50c1VjBz5p2JhUcydE0-Kv0wXmIR9C8X3U,3401
1069
+ wisent/parameters/lm_eval/repair_plan.json,sha256=7fHvz9rYoeJvuuciXJekZq6ok-rI9ndVybbficmukbc,11131
1039
1070
  wisent/parameters/lm_eval/track_progress_not_lm_eval_tasks.json,sha256=6dujwX_RNY1MZFbWMb1MtXYK0CdN1QTC6ppPWpxO_64,3472
1071
+ wisent/parameters/lm_eval/weak_contrastive_pairs.json,sha256=ruwV_qTXWKvWKq6zlufYinzeQ1bz2yZJ6eOk_tAprGU,1202
1072
+ wisent/parameters/lm_eval/working_benchmarks.json,sha256=K1Wk2PU9BnmkVFoAwNJc3RcUODaSb2id6aGNjTSzTS0,3240
1073
+ wisent/parameters/lm_eval/working_benchmarks_categorized.json,sha256=F78AlZhPOWW9Si7Qhq28Yz7OZC1jo0c0oAmeZhVydto,4041
1040
1074
  wisent/parameters/tasks/missing_task_families.json,sha256=h6g_2CYNQIW5SqRdcymDe4miWkGkVoD3yR0eOFZH4aw,65107
1041
1075
  wisent/parameters/tasks/remaining_tasks_to_implement.json,sha256=8_YDVQ439DdvC3daNW7v8y0Zth1eUxum6QzdnKDp6a8,3432
1042
1076
  wisent/parameters/tasks/risks.json,sha256=8U1LAY6Zrv68naYMzVlOrGgTBNL0l-9nC7moa-ljSHI,136
@@ -1046,11 +1080,11 @@ wisent/scripts/run_quality_metrics_sweep.sh,sha256=Lj8euHTGBO3QMdNhpW5zFXWPd4bPP
1046
1080
  wisent/tests/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
1047
1081
  wisent/tests/test_aggregation_geometry.py,sha256=hh8aJ6VyuN_3xA0x9YaoPhu_lMVYKupUxpVyFtEF_18,9277
1048
1082
  wisent/tests/test_all_cli_commands.py,sha256=7KqSZ8FsJEwXbgp0w_cfVtv9jgd-9P52fH3nx3jGF9k,23432
1049
- wisent/tests/test_detector_accuracy.py,sha256=HnmBBwCMZC0bM738o8oCdk4dZmlA8Sr228iLdAbTNio,6484
1083
+ wisent/tests/test_detector_accuracy.py,sha256=9puG4bcjmeHkMsMO62g49Jq5YjfqLA4r9rzsDYXQRU8,6464
1050
1084
  wisent/tests/test_geometry_comprehensive.py,sha256=lUyBfZYudMMaX2-nV2QhtQcchbBaz8bLi6fqmMRQOig,13693
1051
1085
  wisent/tests/test_geometry_exhaustive.py,sha256=yQnUIjNTW8DaC_bwU_rfHThOutLoMGxTHxsbkbqCZNQ,49679
1052
1086
  wisent/tests/test_titan_geometry.py,sha256=lKWsFSCGEHFHD-CW_lz0vfATb6iFij0lrrLosBQJmkc,9710
1053
- wisent/tests/visualize_geometry.py,sha256=dhwyRlqjZfIY9xxJO5rFjLOfdmPyjhg9sjZTcd4qBIY,13861
1087
+ wisent/tests/visualize_geometry.py,sha256=7l_QOoN7Rftmwzu6tn2M6Fug6uniKRSBguwCFeSnTzg,13841
1054
1088
  wisent/tests/examples/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
1055
1089
  wisent/tests/examples/cli/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
1056
1090
  wisent/tests/examples/cli/activations/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -1080,9 +1114,9 @@ wisent/tests/nosense/__init__.py,sha256=sH3x4jRPzFM3YmQkdrwJoz-BdOQ1Bh6F95G5HWyI
1080
1114
  wisent/tests/nosense/base_nosense.py,sha256=a18dBv1378nHly7OCIuk-bCcLnubss3XXDC1ex0zCK8,2633
1081
1115
  wisent/tests/nosense/math500_nosense.py,sha256=My0dHsr4OFOiTxb_VDKmGzpoMyzAtqXlHhA0oPfaG7s,2389
1082
1116
  wisent/tests/nosense/test_robustness.py,sha256=eeKji-_ls6tx7tuXqUO4BXxFRK-giJVihENAJVOvzSs,12546
1083
- wisent-0.7.701.dist-info/licenses/LICENSE,sha256=wy0iaw8b2tyqZAfKHib3lP3PJ9o88FDCg92oUHh3sDQ,1073
1084
- wisent-0.7.701.dist-info/METADATA,sha256=ZO6yC67KziyCpX9sXltKYVciAxX2s5jMm9AYTmyRWt0,2125
1085
- wisent-0.7.701.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
1086
- wisent-0.7.701.dist-info/entry_points.txt,sha256=BM76j3xjtIcVZGk24iDf5w18s6SuqeOpaiAxfZhpnY8,49
1087
- wisent-0.7.701.dist-info/top_level.txt,sha256=2Ts9Iyldnb3auIN2HBBaHPknRy7nSRDm2f6RGzYgr8A,7
1088
- wisent-0.7.701.dist-info/RECORD,,
1117
+ wisent-0.7.901.dist-info/licenses/LICENSE,sha256=wy0iaw8b2tyqZAfKHib3lP3PJ9o88FDCg92oUHh3sDQ,1073
1118
+ wisent-0.7.901.dist-info/METADATA,sha256=gA_lnwxajFyuc_Dr9wFJ7v9cLeQc27KELDeDiDX4ugo,2125
1119
+ wisent-0.7.901.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
1120
+ wisent-0.7.901.dist-info/entry_points.txt,sha256=BM76j3xjtIcVZGk24iDf5w18s6SuqeOpaiAxfZhpnY8,49
1121
+ wisent-0.7.901.dist-info/top_level.txt,sha256=2Ts9Iyldnb3auIN2HBBaHPknRy7nSRDm2f6RGzYgr8A,7
1122
+ wisent-0.7.901.dist-info/RECORD,,