agentinputlab 0.1.0a1__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 (469) hide show
  1. agentinputlab/__init__.py +28 -0
  2. agentinputlab/algorithms/__init__.py +6 -0
  3. agentinputlab/algorithms/acquire/__init__.py +145 -0
  4. agentinputlab/algorithms/acquire/adaptive_rag/__init__.py +43 -0
  5. agentinputlab/algorithms/acquire/adaptive_rag/replay.py +367 -0
  6. agentinputlab/algorithms/acquire/adaptive_rag/runtime.py +1202 -0
  7. agentinputlab/algorithms/acquire/adaptive_rag_common_support.py +542 -0
  8. agentinputlab/algorithms/acquire/hyde.py +296 -0
  9. agentinputlab/algorithms/acquire/hyde_controlled.py +200 -0
  10. agentinputlab/algorithms/acquire/hyde_external.py +148 -0
  11. agentinputlab/algorithms/acquire/ircot.py +28 -0
  12. agentinputlab/algorithms/acquire/ircot_common_support.py +478 -0
  13. agentinputlab/algorithms/acquire/ircot_native.py +355 -0
  14. agentinputlab/algorithms/acquire/ircot_replay.py +327 -0
  15. agentinputlab/algorithms/acquire/qrag/__init__.py +79 -0
  16. agentinputlab/algorithms/acquire/qrag/bridge.py +828 -0
  17. agentinputlab/algorithms/acquire/qrag/common_support.py +891 -0
  18. agentinputlab/algorithms/acquire/qrag/native_axis.py +71 -0
  19. agentinputlab/algorithms/acquire/qrag/reader_adapter.py +139 -0
  20. agentinputlab/algorithms/acquire/r2ag/__init__.py +57 -0
  21. agentinputlab/algorithms/acquire/r2ag/bridge.py +1312 -0
  22. agentinputlab/algorithms/acquire/r2ag/common_support.py +1029 -0
  23. agentinputlab/algorithms/acquire/r2ag/socket_runtime.py +255 -0
  24. agentinputlab/algorithms/clarify/__init__.py +61 -0
  25. agentinputlab/algorithms/clarify/_structured.py +64 -0
  26. agentinputlab/algorithms/clarify/active_task_disambiguation.py +371 -0
  27. agentinputlab/algorithms/clarify/ask_or_assume/__init__.py +17 -0
  28. agentinputlab/algorithms/clarify/ask_or_assume/monitor.py +410 -0
  29. agentinputlab/algorithms/clarify/future_turns.py +379 -0
  30. agentinputlab/algorithms/clarify/udcs.py +255 -0
  31. agentinputlab/algorithms/clarify/udcs_deterministic.py +196 -0
  32. agentinputlab/algorithms/curate/__init__.py +65 -0
  33. agentinputlab/algorithms/curate/acon/__init__.py +27 -0
  34. agentinputlab/algorithms/curate/acon/common.py +284 -0
  35. agentinputlab/algorithms/curate/acon/common_support.py +136 -0
  36. agentinputlab/algorithms/curate/acon/history_bridge.py +309 -0
  37. agentinputlab/algorithms/curate/acon/observation_bridge.py +290 -0
  38. agentinputlab/algorithms/curate/llmlingua/__init__.py +30 -0
  39. agentinputlab/algorithms/curate/llmlingua/adapters.py +537 -0
  40. agentinputlab/algorithms/curate/llmlingua/budget.py +48 -0
  41. agentinputlab/algorithms/curate/llmlingua/controlled.py +175 -0
  42. agentinputlab/algorithms/curate/llmlingua/multilingual.py +105 -0
  43. agentinputlab/algorithms/curate/llmlingua/span_map.py +167 -0
  44. agentinputlab/algorithms/curate/recomp/__init__.py +47 -0
  45. agentinputlab/algorithms/curate/recomp/external.py +1129 -0
  46. agentinputlab/algorithms/curate/recomp/selective.py +304 -0
  47. agentinputlab/algorithms/decompose/__init__.py +107 -0
  48. agentinputlab/algorithms/decompose/decomp/__init__.py +53 -0
  49. agentinputlab/algorithms/decompose/decomp/common_support.py +478 -0
  50. agentinputlab/algorithms/decompose/decomp/planner_component.py +199 -0
  51. agentinputlab/algorithms/decompose/decomp/worker_bridge.py +416 -0
  52. agentinputlab/algorithms/decompose/llm_compiler/__init__.py +35 -0
  53. agentinputlab/algorithms/decompose/llm_compiler/bridge.py +520 -0
  54. agentinputlab/algorithms/decompose/llm_compiler/common_support.py +504 -0
  55. agentinputlab/algorithms/decompose/plan_and_budget/__init__.py +43 -0
  56. agentinputlab/algorithms/decompose/plan_and_budget/cached_soft_budget.py +362 -0
  57. agentinputlab/algorithms/decompose/plan_and_budget/decomposer_component.py +107 -0
  58. agentinputlab/algorithms/decompose/plan_and_budget/full.py +205 -0
  59. agentinputlab/algorithms/decompose/plan_and_budget/hard_cap.py +246 -0
  60. agentinputlab/algorithms/decompose/plan_and_budget/regenerated_soft_budget.py +448 -0
  61. agentinputlab/algorithms/decompose/plan_and_solve/__init__.py +33 -0
  62. agentinputlab/algorithms/decompose/plan_and_solve/plan_component.py +192 -0
  63. agentinputlab/algorithms/decompose/plan_and_solve/ps_prompt.py +127 -0
  64. agentinputlab/algorithms/decompose/self_ask/__init__.py +27 -0
  65. agentinputlab/algorithms/decompose/self_ask/loop.py +362 -0
  66. agentinputlab/algorithms/decompose/self_ask/planner_component.py +117 -0
  67. agentinputlab/algorithms/reformulate/__init__.py +85 -0
  68. agentinputlab/algorithms/reformulate/infocqr/__init__.py +40 -0
  69. agentinputlab/algorithms/reformulate/infocqr/cache_replay.py +362 -0
  70. agentinputlab/algorithms/reformulate/infocqr/external.py +675 -0
  71. agentinputlab/algorithms/reformulate/infocqr/student.py +189 -0
  72. agentinputlab/algorithms/reformulate/query2doc/__init__.py +42 -0
  73. agentinputlab/algorithms/reformulate/query2doc/controlled_live.py +190 -0
  74. agentinputlab/algorithms/reformulate/query2doc/models.py +95 -0
  75. agentinputlab/algorithms/reformulate/query2doc/official_cached.py +335 -0
  76. agentinputlab/algorithms/reformulate/query2doc/querygym_bridge.py +223 -0
  77. agentinputlab/algorithms/reformulate/query2doc/querygym_runtime.py +326 -0
  78. agentinputlab/algorithms/reformulate/reformer/__init__.py +23 -0
  79. agentinputlab/algorithms/reformulate/reformer/external.py +417 -0
  80. agentinputlab/algorithms/reformulate/rrr.py +696 -0
  81. agentinputlab/annotation/__init__.py +5 -0
  82. agentinputlab/annotation/independent.py +428 -0
  83. agentinputlab/annotation/server.py +233 -0
  84. agentinputlab/annotation/static/app.js +121 -0
  85. agentinputlab/annotation/static/index.html +60 -0
  86. agentinputlab/annotation/static/review.css +2 -0
  87. agentinputlab/annotation/static/review.html +10 -0
  88. agentinputlab/annotation/static/review.js +11 -0
  89. agentinputlab/annotation/static/styles.css +55 -0
  90. agentinputlab/annotation/store.py +444 -0
  91. agentinputlab/api.py +368 -0
  92. agentinputlab/artifacts/__init__.py +15 -0
  93. agentinputlab/artifacts/json_file.py +104 -0
  94. agentinputlab/artifacts/manifest.py +98 -0
  95. agentinputlab/artifacts/store.py +157 -0
  96. agentinputlab/backends/__init__.py +98 -0
  97. agentinputlab/backends/embedding/__init__.py +25 -0
  98. agentinputlab/backends/embedding/base.py +145 -0
  99. agentinputlab/backends/embedding/callable.py +75 -0
  100. agentinputlab/backends/embedding/replay.py +129 -0
  101. agentinputlab/backends/llm/__init__.py +31 -0
  102. agentinputlab/backends/llm/base.py +149 -0
  103. agentinputlab/backends/llm/callable.py +108 -0
  104. agentinputlab/backends/llm/decomp_completion.py +88 -0
  105. agentinputlab/backends/llm/final_content_stop.py +72 -0
  106. agentinputlab/backends/llm/gpt_oss_final_channel.py +84 -0
  107. agentinputlab/backends/llm/openai_compatible.py +332 -0
  108. agentinputlab/backends/llm/transformers.py +171 -0
  109. agentinputlab/backends/retrieval/__init__.py +35 -0
  110. agentinputlab/backends/retrieval/base.py +50 -0
  111. agentinputlab/backends/retrieval/dense.py +777 -0
  112. agentinputlab/backends/retrieval/memory.py +119 -0
  113. agentinputlab/backends/retrieval/sqlite_fts.py +155 -0
  114. agentinputlab/backends/user/__init__.py +15 -0
  115. agentinputlab/backends/user/base.py +66 -0
  116. agentinputlab/backends/user/callable.py +54 -0
  117. agentinputlab/backends/user/replay.py +89 -0
  118. agentinputlab/backends/vector.py +38 -0
  119. agentinputlab/benchmarks/__init__.py +246 -0
  120. agentinputlab/benchmarks/acquisition.py +106 -0
  121. agentinputlab/benchmarks/ale.py +540 -0
  122. agentinputlab/benchmarks/appworld_ul.py +669 -0
  123. agentinputlab/benchmarks/assets.py +226 -0
  124. agentinputlab/benchmarks/base.py +68 -0
  125. agentinputlab/benchmarks/beir_archive.py +235 -0
  126. agentinputlab/benchmarks/beir_sample.py +269 -0
  127. agentinputlab/benchmarks/bfcl_native.py +245 -0
  128. agentinputlab/benchmarks/bfcl_v3.py +607 -0
  129. agentinputlab/benchmarks/cache.py +218 -0
  130. agentinputlab/benchmarks/clariq.py +203 -0
  131. agentinputlab/benchmarks/data/registry.json +321 -0
  132. agentinputlab/benchmarks/data/smoke_assets.json +511 -0
  133. agentinputlab/benchmarks/data/statebench_v0_8_1_task_ids.json +475 -0
  134. agentinputlab/benchmarks/data/toolsandbox_165848b_scenario_names.json +1 -0
  135. agentinputlab/benchmarks/execution.py +1111 -0
  136. agentinputlab/benchmarks/interactive.py +257 -0
  137. agentinputlab/benchmarks/ir_query.py +204 -0
  138. agentinputlab/benchmarks/longbench.py +203 -0
  139. agentinputlab/benchmarks/paired_replay.py +479 -0
  140. agentinputlab/benchmarks/preprocess.py +2313 -0
  141. agentinputlab/benchmarks/production.py +236 -0
  142. agentinputlab/benchmarks/qa.py +963 -0
  143. agentinputlab/benchmarks/registry.py +110 -0
  144. agentinputlab/benchmarks/retrieval.py +459 -0
  145. agentinputlab/benchmarks/runtime.py +509 -0
  146. agentinputlab/benchmarks/statebench.py +938 -0
  147. agentinputlab/benchmarks/tau3.py +725 -0
  148. agentinputlab/benchmarks/toolsandbox.py +626 -0
  149. agentinputlab/benchmarks/trec_dl_sample.py +263 -0
  150. agentinputlab/campaign.py +654 -0
  151. agentinputlab/campaign_runner.py +953 -0
  152. agentinputlab/cli/__init__.py +6 -0
  153. agentinputlab/cli/main.py +933 -0
  154. agentinputlab/connectors/__init__.py +46 -0
  155. agentinputlab/connectors/knowledge/__init__.py +4 -0
  156. agentinputlab/connectors/tools/__init__.py +30 -0
  157. agentinputlab/connectors/tools/base.py +57 -0
  158. agentinputlab/connectors/tools/mcp.py +222 -0
  159. agentinputlab/connectors/tools/memory.py +41 -0
  160. agentinputlab/connectors/tools/openapi.py +695 -0
  161. agentinputlab/connectors/tools/schema.py +102 -0
  162. agentinputlab/contracts/__init__.py +91 -0
  163. agentinputlab/contracts/availability.py +24 -0
  164. agentinputlab/contracts/dependency.py +58 -0
  165. agentinputlab/contracts/enums.py +82 -0
  166. agentinputlab/contracts/errors.py +63 -0
  167. agentinputlab/contracts/evidence.py +43 -0
  168. agentinputlab/contracts/intent.py +110 -0
  169. agentinputlab/contracts/lifecycle.py +164 -0
  170. agentinputlab/contracts/patch.py +167 -0
  171. agentinputlab/contracts/serialization.py +93 -0
  172. agentinputlab/contracts/state.py +127 -0
  173. agentinputlab/contracts/trace.py +59 -0
  174. agentinputlab/contracts/worker.py +42 -0
  175. agentinputlab/core/__init__.py +7 -0
  176. agentinputlab/data/__init__.py +32 -0
  177. agentinputlab/data/checksums.py +9 -0
  178. agentinputlab/data/downloader.py +235 -0
  179. agentinputlab/data/http.py +171 -0
  180. agentinputlab/data/index.py +21 -0
  181. agentinputlab/data/overlap.py +28 -0
  182. agentinputlab/data/preprocess.py +32 -0
  183. agentinputlab/data/storage.py +57 -0
  184. agentinputlab/data/trec_sample.py +427 -0
  185. agentinputlab/evaluation/__init__.py +5 -0
  186. agentinputlab/evaluation/outcomes.py +32 -0
  187. agentinputlab/experiments/__init__.py +278 -0
  188. agentinputlab/experiments/analysis.py +981 -0
  189. agentinputlab/experiments/artifacts.py +687 -0
  190. agentinputlab/experiments/automatic_measurement.py +678 -0
  191. agentinputlab/experiments/auxiliary_matrix.py +203 -0
  192. agentinputlab/experiments/bindings.py +430 -0
  193. agentinputlab/experiments/blueprints.py +147 -0
  194. agentinputlab/experiments/builders/__init__.py +45 -0
  195. agentinputlab/experiments/builders/executable_external_payload.py +423 -0
  196. agentinputlab/experiments/builders/external_process_arena.py +1018 -0
  197. agentinputlab/experiments/builders/registry_intervention.py +463 -0
  198. agentinputlab/experiments/cli.py +158 -0
  199. agentinputlab/experiments/controls.py +340 -0
  200. agentinputlab/experiments/driver_guard.py +205 -0
  201. agentinputlab/experiments/drivers/__init__.py +89 -0
  202. agentinputlab/experiments/drivers/external_arena.py +2908 -0
  203. agentinputlab/experiments/drivers/method_process.py +662 -0
  204. agentinputlab/experiments/drivers/parity_process.py +1484 -0
  205. agentinputlab/experiments/drivers/static_production.py +1574 -0
  206. agentinputlab/experiments/executor.py +448 -0
  207. agentinputlab/experiments/loader.py +605 -0
  208. agentinputlab/experiments/measurement.py +675 -0
  209. agentinputlab/experiments/paper_analysis.py +1818 -0
  210. agentinputlab/experiments/parity_identity.py +605 -0
  211. agentinputlab/experiments/plan.py +277 -0
  212. agentinputlab/experiments/production_builders.py +3739 -0
  213. agentinputlab/experiments/readiness.py +432 -0
  214. agentinputlab/experiments/schema.py +514 -0
  215. agentinputlab/experiments/secrets.py +177 -0
  216. agentinputlab/experiments/stateful_analysis.py +1738 -0
  217. agentinputlab/experiments/transfer_analysis.py +1316 -0
  218. agentinputlab/experiments/unseen_analysis.py +1025 -0
  219. agentinputlab/integrations/__init__.py +1 -0
  220. agentinputlab/integrations/action_space/__init__.py +1 -0
  221. agentinputlab/integrations/external_workers/__init__.py +76 -0
  222. agentinputlab/integrations/external_workers/ace.py +431 -0
  223. agentinputlab/integrations/external_workers/base.py +181 -0
  224. agentinputlab/integrations/external_workers/decomp_live.py +249 -0
  225. agentinputlab/integrations/external_workers/decomp_planner.py +30 -0
  226. agentinputlab/integrations/external_workers/infocqr_student.py +280 -0
  227. agentinputlab/integrations/external_workers/ircot.py +576 -0
  228. agentinputlab/integrations/external_workers/llmcompiler_live.py +360 -0
  229. agentinputlab/integrations/external_workers/longllmlingua.py +162 -0
  230. agentinputlab/integrations/external_workers/method.py +314 -0
  231. agentinputlab/integrations/external_workers/process.py +678 -0
  232. agentinputlab/integrations/external_workers/supervisor.py +1071 -0
  233. agentinputlab/integrations/full_architecture/__init__.py +1 -0
  234. agentinputlab/integrations/full_architecture/hyde.py +212 -0
  235. agentinputlab/integrations/full_architecture/plan_and_budget.py +188 -0
  236. agentinputlab/integrations/live_history.py +222 -0
  237. agentinputlab/integrations/prompt_optimization/__init__.py +1 -0
  238. agentinputlab/integrations/training/__init__.py +1 -0
  239. agentinputlab/metrics/__init__.py +50 -0
  240. agentinputlab/metrics/catalog.py +313 -0
  241. agentinputlab/metrics/harm.py +95 -0
  242. agentinputlab/metrics/paired.py +40 -0
  243. agentinputlab/metrics/qa.py +76 -0
  244. agentinputlab/metrics/text.py +33 -0
  245. agentinputlab/metrics/third_party.py +161 -0
  246. agentinputlab/models/__init__.py +4 -0
  247. agentinputlab/observability/__init__.py +56 -0
  248. agentinputlab/observability/budget.py +316 -0
  249. agentinputlab/observability/instrumentation.py +754 -0
  250. agentinputlab/observability/models.py +403 -0
  251. agentinputlab/observability/monitor.py +380 -0
  252. agentinputlab/optimizers/__init__.py +1 -0
  253. agentinputlab/optimizers/clarify/__init__.py +12 -0
  254. agentinputlab/optimizers/context/__init__.py +22 -0
  255. agentinputlab/optimizers/decompose/__init__.py +64 -0
  256. agentinputlab/optimizers/evidence/__init__.py +22 -0
  257. agentinputlab/optimizers/rewrite/__init__.py +30 -0
  258. agentinputlab/parity/__init__.py +5 -0
  259. agentinputlab/parity/base.py +206 -0
  260. agentinputlab/production_campaign.py +763 -0
  261. agentinputlab/py.typed +1 -0
  262. agentinputlab/recipes/__init__.py +333 -0
  263. agentinputlab/recipes/ace.py +418 -0
  264. agentinputlab/recipes/ace_bfcl.py +273 -0
  265. agentinputlab/recipes/ace_common_support.py +169 -0
  266. agentinputlab/recipes/ace_native_plan.py +322 -0
  267. agentinputlab/recipes/ace_online.py +711 -0
  268. agentinputlab/recipes/ace_retrieval.py +330 -0
  269. agentinputlab/recipes/ace_statebench.py +382 -0
  270. agentinputlab/recipes/ace_tau.py +463 -0
  271. agentinputlab/recipes/ace_toolsandbox.py +432 -0
  272. agentinputlab/recipes/acon_guideline.py +22 -0
  273. agentinputlab/recipes/adaptive_rag.py +822 -0
  274. agentinputlab/recipes/llmlingua2_inference.py +146 -0
  275. agentinputlab/recipes/llmlingua2_training.py +796 -0
  276. agentinputlab/recipes/mce.py +1486 -0
  277. agentinputlab/recipes/mce_final.py +335 -0
  278. agentinputlab/recipes/mce_text_fit.py +499 -0
  279. agentinputlab/recipes/mce_text_only.py +298 -0
  280. agentinputlab/recipes/qrag.py +729 -0
  281. agentinputlab/recipes/qrag_inference.py +364 -0
  282. agentinputlab/recipes/r2ag_features.py +236 -0
  283. agentinputlab/recipes/r2ag_inference.py +503 -0
  284. agentinputlab/recipes/r2ag_training.py +948 -0
  285. agentinputlab/recipes/recomp_abstractive.py +592 -0
  286. agentinputlab/recipes/recomp_extractive.py +525 -0
  287. agentinputlab/recipes/reformer.py +17 -0
  288. agentinputlab/recipes/rqrag.py +1516 -0
  289. agentinputlab/recipes/rqrag_common_support.py +193 -0
  290. agentinputlab/recipes/rqrag_completion.py +147 -0
  291. agentinputlab/recipes/skillopt.py +462 -0
  292. agentinputlab/recipes/skillopt_common_support.py +154 -0
  293. agentinputlab/recipes/skillopt_environment.py +165 -0
  294. agentinputlab/recipes/skillopt_environment_cli.py +5 -0
  295. agentinputlab/recipes/skillopt_retrieval.py +188 -0
  296. agentinputlab/recipes/smartsearch.py +897 -0
  297. agentinputlab/recipes/smartsearch_curriculum.py +283 -0
  298. agentinputlab/recipes/smartsearch_grpo_contracts.py +58 -0
  299. agentinputlab/recipes/smartsearch_grpo_data.py +314 -0
  300. agentinputlab/recipes/smartsearch_grpo_meter.py +324 -0
  301. agentinputlab/recipes/smartsearch_grpo_prefix.py +144 -0
  302. agentinputlab/recipes/smartsearch_grpo_rollout.py +119 -0
  303. agentinputlab/recipes/smartsearch_grpo_tools.py +364 -0
  304. agentinputlab/recipes/smartsearch_grpo_training.py +305 -0
  305. agentinputlab/recipes/smartsearch_grpo_transport.py +248 -0
  306. agentinputlab/recipes/smartsearch_grpo_worker.py +364 -0
  307. agentinputlab/recipes/smartsearch_inference.py +524 -0
  308. agentinputlab/recipes/smartsearch_native_data.py +172 -0
  309. agentinputlab/recipes/smartsearch_policy.py +616 -0
  310. agentinputlab/recipes/smartsearch_policy_worker.py +9 -0
  311. agentinputlab/recipes/smartsearch_sampling.py +232 -0
  312. agentinputlab/recipes/smartsearch_teacher.py +203 -0
  313. agentinputlab/recipes/smartsearch_training.py +265 -0
  314. agentinputlab/registry/__init__.py +46 -0
  315. agentinputlab/registry/beta.py +727 -0
  316. agentinputlab/registry/data/algorithms.json +1959 -0
  317. agentinputlab/registry/data/auxiliary_method_profiles.json +34 -0
  318. agentinputlab/registry/loader.py +100 -0
  319. agentinputlab/registry/result_manifest.py +184 -0
  320. agentinputlab/registry/runtime.py +195 -0
  321. agentinputlab/registry/schema.py +449 -0
  322. agentinputlab/runner/__init__.py +30 -0
  323. agentinputlab/runner/balance.py +248 -0
  324. agentinputlab/runner/cache.py +135 -0
  325. agentinputlab/runner/composite.py +130 -0
  326. agentinputlab/runner/engine.py +285 -0
  327. agentinputlab/runner/runtime_guard.py +268 -0
  328. agentinputlab/runner/worker.py +80 -0
  329. agentinputlab/runtime_bootstrap.py +288 -0
  330. agentinputlab/selectors/__init__.py +196 -0
  331. agentinputlab/selectors/advanced.py +1514 -0
  332. agentinputlab/selectors/artifacts.py +1087 -0
  333. agentinputlab/selectors/base.py +112 -0
  334. agentinputlab/selectors/campaigns.py +1893 -0
  335. agentinputlab/selectors/catalog.py +86 -0
  336. agentinputlab/selectors/discovery.py +1860 -0
  337. agentinputlab/selectors/evaluation.py +1348 -0
  338. agentinputlab/selectors/features.py +383 -0
  339. agentinputlab/selectors/learned.py +502 -0
  340. agentinputlab/selectors/no_op.py +28 -0
  341. agentinputlab/selectors/outcome_table.py +475 -0
  342. agentinputlab/selectors/producer.py +853 -0
  343. agentinputlab/selectors/rules.py +109 -0
  344. agentinputlab/selectors/runtime.py +420 -0
  345. agentinputlab/studies/__init__.py +16 -0
  346. agentinputlab/studies/design.py +165 -0
  347. agentinputlab/studies/paired.py +322 -0
  348. agentinputlab/studio/__init__.py +6 -0
  349. agentinputlab/studio/campaigns.py +819 -0
  350. agentinputlab/studio/engine.py +1408 -0
  351. agentinputlab/studio/model_backends.py +425 -0
  352. agentinputlab/studio/prepared.py +410 -0
  353. agentinputlab/studio/server.py +613 -0
  354. agentinputlab/studio/static/app.js +397 -0
  355. agentinputlab/studio/static/campaigns.html +175 -0
  356. agentinputlab/studio/static/campaigns.js +426 -0
  357. agentinputlab/studio/static/index.html +192 -0
  358. agentinputlab/studio/static/result.html +103 -0
  359. agentinputlab/studio/static/result.js +273 -0
  360. agentinputlab/studio/static/styles.css +467 -0
  361. agentinputlab-0.1.0a1.data/data/share/agentinputlab/CITATION.cff +22 -0
  362. agentinputlab-0.1.0a1.dist-info/METADATA +392 -0
  363. agentinputlab-0.1.0a1.dist-info/RECORD +469 -0
  364. agentinputlab-0.1.0a1.dist-info/WHEEL +5 -0
  365. agentinputlab-0.1.0a1.dist-info/entry_points.txt +6 -0
  366. agentinputlab-0.1.0a1.dist-info/licenses/LICENSE +21 -0
  367. agentinputlab-0.1.0a1.dist-info/licenses/NOTICE +31 -0
  368. agentinputlab-0.1.0a1.dist-info/licenses/THIRD_PARTY_NOTICES.md +94 -0
  369. agentinputlab-0.1.0a1.dist-info/top_level.txt +2 -0
  370. workers/README.md +94 -0
  371. workers/ace/README.md +261 -0
  372. workers/ace/__init__.py +1 -0
  373. workers/ace/compatibility_manifest.json +39 -0
  374. workers/ace/driver.py +1682 -0
  375. workers/acon/README.md +215 -0
  376. workers/acon/__init__.py +5 -0
  377. workers/acon/compatibility_manifest.json +93 -0
  378. workers/acon/driver.py +1859 -0
  379. workers/acon/fitted.py +395 -0
  380. workers/acon/guideline.py +470 -0
  381. workers/acon/guideline_source_manifest.json +20 -0
  382. workers/adaptive_rag/README.md +96 -0
  383. workers/adaptive_rag/__init__.py +15 -0
  384. workers/adaptive_rag/compatibility_manifest.json +61 -0
  385. workers/adaptive_rag/driver.py +1470 -0
  386. workers/adaptive_rag/fixture_runtime.py +163 -0
  387. workers/adaptive_rag/runtime.py +252 -0
  388. workers/ask_or_assume/README.md +124 -0
  389. workers/ask_or_assume/__init__.py +1 -0
  390. workers/ask_or_assume/compatibility_manifest.json +36 -0
  391. workers/ask_or_assume/driver.py +1227 -0
  392. workers/decomp/README.md +133 -0
  393. workers/decomp/UPSTREAM_LICENSE.Apache-2.0.txt +201 -0
  394. workers/decomp/__init__.py +5 -0
  395. workers/decomp/compatibility_manifest.json +205 -0
  396. workers/decomp/driver.py +1015 -0
  397. workers/decomp/live.py +137 -0
  398. workers/decomp/runtime.py +231 -0
  399. workers/infocqr/README.md +106 -0
  400. workers/infocqr/__init__.py +1 -0
  401. workers/infocqr/compatibility_manifest.json +84 -0
  402. workers/infocqr/driver.py +1766 -0
  403. workers/ircot/README.md +167 -0
  404. workers/ircot/__init__.py +7 -0
  405. workers/ircot/compatibility_manifest.json +121 -0
  406. workers/ircot/controlled.py +457 -0
  407. workers/ircot/driver.py +1373 -0
  408. workers/ircot/fixture_bindings.json +39 -0
  409. workers/ircot/fixture_config.json +48 -0
  410. workers/ircot/fixture_prompt.txt +2 -0
  411. workers/ircot/native_qa.py +398 -0
  412. workers/ircot/qa_compatibility_manifest.json +57 -0
  413. workers/ircot/runtime.py +97 -0
  414. workers/llmcompiler/README.md +160 -0
  415. workers/llmcompiler/__init__.py +1 -0
  416. workers/llmcompiler/compatibility_manifest.json +45 -0
  417. workers/llmcompiler/driver.py +1514 -0
  418. workers/llmcompiler/export_prompts.py +74 -0
  419. workers/llmcompiler/live.py +174 -0
  420. workers/llmcompiler/patch_manifest.json +147 -0
  421. workers/mce/README.md +198 -0
  422. workers/mce/__init__.py +29 -0
  423. workers/mce/compatibility_manifest.json +115 -0
  424. workers/mce/driver.py +2585 -0
  425. workers/mce/final_eval.py +173 -0
  426. workers/official_arenas/README.md +25 -0
  427. workers/official_arenas/__init__.py +1 -0
  428. workers/official_arenas/capture_records.py +110 -0
  429. workers/official_arenas/driver.py +619 -0
  430. workers/official_arenas/protocol.py +334 -0
  431. workers/official_arenas/providers/__init__.py +1 -0
  432. workers/official_arenas/providers/_live_history.py +228 -0
  433. workers/official_arenas/providers/_message_compat.py +77 -0
  434. workers/official_arenas/providers/appworld_ul.py +196 -0
  435. workers/official_arenas/providers/fixture.py +112 -0
  436. workers/official_arenas/providers/statebench.py +960 -0
  437. workers/official_arenas/providers/tau3.py +1672 -0
  438. workers/official_arenas/providers/tau_retail.py +586 -0
  439. workers/official_arenas/providers/toolsandbox.py +1321 -0
  440. workers/official_arenas/providers/tooltalk.py +699 -0
  441. workers/official_arenas/retail_capture.py +188 -0
  442. workers/official_arenas/retail_driver.py +275 -0
  443. workers/qrag/README.md +172 -0
  444. workers/qrag/__init__.py +1 -0
  445. workers/qrag/compatibility_manifest.json +64 -0
  446. workers/qrag/driver.py +673 -0
  447. workers/r2ag/README.md +154 -0
  448. workers/r2ag/__init__.py +5 -0
  449. workers/r2ag/compatibility_manifest.json +55 -0
  450. workers/r2ag/driver.py +1287 -0
  451. workers/r2ag/feature_schema.json +8 -0
  452. workers/r2ag/runtime.py +589 -0
  453. workers/recomp/README.md +120 -0
  454. workers/recomp/__init__.py +1 -0
  455. workers/recomp/compatibility_manifest.json +43 -0
  456. workers/recomp/driver.py +1653 -0
  457. workers/reformer/README.md +125 -0
  458. workers/reformer/__init__.py +1 -0
  459. workers/reformer/compatibility_manifest.json +68 -0
  460. workers/reformer/driver.py +1459 -0
  461. workers/reformer/induction.py +457 -0
  462. workers/rrr/README.md +126 -0
  463. workers/rrr/__init__.py +1 -0
  464. workers/rrr/compatibility_manifest.json +57 -0
  465. workers/rrr/driver.py +1133 -0
  466. workers/smartsearch/README.md +331 -0
  467. workers/smartsearch/__init__.py +1 -0
  468. workers/smartsearch/compatibility_manifest.json +41 -0
  469. workers/smartsearch/driver.py +1643 -0
@@ -0,0 +1,511 @@
1
+ {
2
+ "schema_version": 1,
3
+ "asset_count": 16,
4
+ "assets": [
5
+ {
6
+ "asset_id": "beir.arguana.official",
7
+ "cell_id": "beir.arguana",
8
+ "source_url": "https://public.ukp.informatik.tu-darmstadt.de/thakur/BEIR/datasets/arguana.zip",
9
+ "source_revision": "2021-01-23",
10
+ "sha256": "cfdf79adce27a401b3cd3ea267903134dbfab2c6afeb95d7fe5724a00bf7557b",
11
+ "size_bytes": 3773617,
12
+ "license_id": "CC-BY-SA-4.0",
13
+ "redistribution": "downloader_only",
14
+ "status": "verified_smoke",
15
+ "processor": "beir_zip_v1",
16
+ "expected_counts": {
17
+ "corpus": 8674,
18
+ "gold": 1406,
19
+ "inputs": 1406,
20
+ "orphan_qrels": 5,
21
+ "orphan_queries": 5,
22
+ "qrels": 1406
23
+ },
24
+ "prepared_sha256": {
25
+ "corpus.jsonl": "d7916d8e3f9fb56433202043c8fded6c1caaa116d1f0a6765614febaea220618",
26
+ "gold.jsonl": "9215b7caaa8822888753f28aa20203ef6dfb5ef33f92a1cc44ce42266da8cee8",
27
+ "inputs.jsonl": "9c72c6856de16d0d5151e88e2a42fef82175e828eae3b356ba5c759c530b577a"
28
+ },
29
+ "allowed_hosts": [
30
+ "public.ukp.informatik.tu-darmstadt.de"
31
+ ],
32
+ "provenance": {
33
+ "dataset_home": "https://www.tira.io/task/argument-search/",
34
+ "download_index": "https://public.ukp.informatik.tu-darmstadt.de/thakur/BEIR/datasets/",
35
+ "source_repository": "https://github.com/beir-cellar/beir"
36
+ },
37
+ "note": "Full compact BEIR retrieval cell; source bytes are never included in the package. The pinned archive contains five qrels whose relevant corpus IDs are absent; they are preserved and explicitly counted rather than silently dropped."
38
+ },
39
+ {
40
+ "asset_id": "bfcl.v3.simple_python.v1_3",
41
+ "cell_id": "bfcl.v3.simple_python",
42
+ "source_url": "https://github.com/ShishirPatil/gorilla/archive/refs/tags/v1.3.tar.gz",
43
+ "source_revision": "ea13468e4423454d0c213704fb87cf7cb3990433",
44
+ "sha256": "57854157e82aed5723a1458d2a4ac426e3256959ded4a20ddc4d627c22edb7fe",
45
+ "size_bytes": 44552456,
46
+ "license_id": "Apache-2.0",
47
+ "redistribution": "downloader_only",
48
+ "status": "verified_smoke",
49
+ "processor": "bfcl_v3_simple_tar_v1",
50
+ "expected_counts": {
51
+ "audit": 1,
52
+ "functions": 400,
53
+ "gold": 400,
54
+ "ground_truth_candidates": 400,
55
+ "inputs": 400,
56
+ "source_files": 8
57
+ },
58
+ "prepared_sha256": {
59
+ "audit.jsonl": "55f7039db2bf049b99ff31ff0c1edd0356b6b074a77af6f39a8527087105d24b",
60
+ "gold.jsonl": "cebbd2664755ef02e7a9f8c557be124b4004bce02de829dbe4dc515f9297f352",
61
+ "inputs.jsonl": "c9fab55c3ad0309265d238cff6b736fdad047fba0d1476e2b56c4c241be0fc18"
62
+ },
63
+ "allowed_hosts": [
64
+ "codeload.github.com",
65
+ "github.com"
66
+ ],
67
+ "provenance": {
68
+ "official_release": "https://github.com/ShishirPatil/gorilla/releases/tag/v1.3",
69
+ "source_repository": "https://github.com/ShishirPatil/gorilla",
70
+ "v3_announcement": "https://gorilla.cs.berkeley.edu/blogs/8_berkeley_function_calling_leaderboard.html"
71
+ },
72
+ "note": "Official v1.3 source archive; only the 400-case BFCL_v3_simple Python non-live cell is normalized. Possible answers remain evaluator-only. Source bytes, model outputs, and result claims are never included in the package."
73
+ },
74
+ {
75
+ "asset_id": "bright.earth_science.examples",
76
+ "cell_id": "bright.earth_science",
77
+ "source_url": "https://huggingface.co/datasets/xlangai/BRIGHT/resolve/1c9ff4a4090306cd8d9412b108023f1f10942864/examples/earth_science-00000-of-00001.parquet",
78
+ "source_revision": "1c9ff4a4090306cd8d9412b108023f1f10942864",
79
+ "sha256": "5d29f108701111984eb91c93d3e340a784a99df47dc11f43783d1b994010a91d",
80
+ "size_bytes": 184093,
81
+ "license_id": "CC-BY-4.0",
82
+ "redistribution": "downloader_only",
83
+ "status": "verified_smoke",
84
+ "processor": "bright_examples_parquet_v1",
85
+ "expected_counts": {
86
+ "excluded_documents": 0,
87
+ "gold": 116,
88
+ "inputs": 116,
89
+ "qrels": 585
90
+ },
91
+ "prepared_sha256": {
92
+ "gold.jsonl": "512eaa4220ed269772aab8f97613d16d7b1c531e4c93759921225c9ca6f70fb8",
93
+ "inputs.jsonl": "dec7334497de6ad63940dfda490f5f1cf552e0208c682d226887548be4f9c9b2"
94
+ },
95
+ "allowed_hosts": [
96
+ "cas-bridge.xethub.hf.co",
97
+ "cdn-lfs.huggingface.co",
98
+ "huggingface.co",
99
+ "us.aws.cdn.hf.co"
100
+ ],
101
+ "provenance": {
102
+ "dataset_card": "https://huggingface.co/datasets/xlangai/BRIGHT",
103
+ "dataset_home": "https://brightbenchmark.github.io/",
104
+ "source_repository": "https://github.com/xlang-ai/BRIGHT"
105
+ },
106
+ "note": "Official Earth Science query/qrel component. Pair with bright.earth_science.corpus; reasoning, answer, long qrels, and exclusions remain evaluator-only."
107
+ },
108
+ {
109
+ "asset_id": "bright.earth_science.corpus",
110
+ "cell_id": "bright.earth_science",
111
+ "source_url": "https://huggingface.co/datasets/xlangai/BRIGHT/resolve/ba248cfa7b4a12da78640f81a3dd064632e25fa6/documents/earth_science-00000-of-00001.parquet",
112
+ "source_revision": "ba248cfa7b4a12da78640f81a3dd064632e25fa6",
113
+ "sha256": "abcb2cd130d7d333b364bf5c89b7ff3829e0f88eb6a2de8232c3df1173eeb8a2",
114
+ "size_bytes": 23084671,
115
+ "license_id": "CC-BY-4.0",
116
+ "redistribution": "downloader_only",
117
+ "status": "verified_smoke",
118
+ "processor": "bright_corpus_parquet_v1",
119
+ "expected_counts": {
120
+ "corpus": 121249
121
+ },
122
+ "prepared_sha256": {
123
+ "corpus.jsonl": "be5947fb747d4c0f3506bb71b013defb4b3a96d49d75fc5f24df17ecec583e69"
124
+ },
125
+ "allowed_hosts": [
126
+ "cas-bridge.xethub.hf.co",
127
+ "cdn-lfs.huggingface.co",
128
+ "huggingface.co",
129
+ "us.aws.cdn.hf.co"
130
+ ],
131
+ "provenance": {
132
+ "dataset_card": "https://huggingface.co/datasets/xlangai/BRIGHT",
133
+ "dataset_home": "https://brightbenchmark.github.io/",
134
+ "source_repository": "https://github.com/xlang-ai/BRIGHT"
135
+ },
136
+ "note": "Official Earth Science corpus component for bright.earth_science.examples; kept separately to preserve upstream layout and component hashes."
137
+ },
138
+ {
139
+ "asset_id": "bright.economics.examples",
140
+ "cell_id": "bright.economics",
141
+ "source_url": "https://huggingface.co/datasets/xlangai/BRIGHT/resolve/1c9ff4a4090306cd8d9412b108023f1f10942864/examples/economics-00000-of-00001.parquet",
142
+ "source_revision": "1c9ff4a4090306cd8d9412b108023f1f10942864",
143
+ "sha256": "2a79f0f3a881c7c03a258cf8ef8ac2db1ca9080963252d9a020bb45a264aa037",
144
+ "size_bytes": 219518,
145
+ "license_id": "CC-BY-4.0",
146
+ "redistribution": "downloader_only",
147
+ "status": "verified_smoke",
148
+ "processor": "bright_examples_parquet_v1",
149
+ "expected_counts": {
150
+ "excluded_documents": 0,
151
+ "gold": 103,
152
+ "inputs": 103,
153
+ "qrels": 800
154
+ },
155
+ "prepared_sha256": {
156
+ "gold.jsonl": "e001474d30c67440da776fc5d0be893cdf183b73be93128343b4174ecf324951",
157
+ "inputs.jsonl": "42bf64a62f5628eb0e3e93e74fc1d36844314018ba19ced13351aad06ecfff2f"
158
+ },
159
+ "allowed_hosts": [
160
+ "cas-bridge.xethub.hf.co",
161
+ "cdn-lfs.huggingface.co",
162
+ "huggingface.co",
163
+ "us.aws.cdn.hf.co"
164
+ ],
165
+ "provenance": {
166
+ "dataset_card": "https://huggingface.co/datasets/xlangai/BRIGHT",
167
+ "dataset_home": "https://brightbenchmark.github.io/",
168
+ "source_repository": "https://github.com/xlang-ai/BRIGHT"
169
+ },
170
+ "note": "Official Economics query/qrel component. Pair with bright.economics.corpus; answer and reasoning annotations remain evaluator-only."
171
+ },
172
+ {
173
+ "asset_id": "bright.economics.corpus",
174
+ "cell_id": "bright.economics",
175
+ "source_url": "https://huggingface.co/datasets/xlangai/BRIGHT/resolve/ba248cfa7b4a12da78640f81a3dd064632e25fa6/documents/economics-00000-of-00001.parquet",
176
+ "source_revision": "ba248cfa7b4a12da78640f81a3dd064632e25fa6",
177
+ "sha256": "f3ba8a0fbc9a9aed07b4970cc686e32cfefcd06d6922402587adf871f006394c",
178
+ "size_bytes": 10969621,
179
+ "license_id": "CC-BY-4.0",
180
+ "redistribution": "downloader_only",
181
+ "status": "verified_smoke",
182
+ "processor": "bright_corpus_parquet_v1",
183
+ "expected_counts": {
184
+ "corpus": 50220
185
+ },
186
+ "prepared_sha256": {
187
+ "corpus.jsonl": "1836a15aeae9a6686b504ce7b647240e2c1a32a3c6b5a21ed3c4444614708c1f"
188
+ },
189
+ "allowed_hosts": [
190
+ "cas-bridge.xethub.hf.co",
191
+ "cdn-lfs.huggingface.co",
192
+ "huggingface.co",
193
+ "us.aws.cdn.hf.co"
194
+ ],
195
+ "provenance": {
196
+ "dataset_card": "https://huggingface.co/datasets/xlangai/BRIGHT",
197
+ "dataset_home": "https://brightbenchmark.github.io/",
198
+ "source_repository": "https://github.com/xlang-ai/BRIGHT"
199
+ },
200
+ "note": "Official Economics corpus component for bright.economics.examples; kept separately to preserve upstream layout and component hashes."
201
+ },
202
+ {
203
+ "asset_id": "bright.robotics.examples",
204
+ "cell_id": "bright.robotics",
205
+ "source_url": "https://huggingface.co/datasets/xlangai/BRIGHT/resolve/1c9ff4a4090306cd8d9412b108023f1f10942864/examples/robotics-00000-of-00001.parquet",
206
+ "source_revision": "1c9ff4a4090306cd8d9412b108023f1f10942864",
207
+ "sha256": "621484c87c9ebae12f81e32a0a8c5d085af4b95cbe1b575ab40ae4b659adb53a",
208
+ "size_bytes": 178820,
209
+ "license_id": "CC-BY-4.0",
210
+ "redistribution": "downloader_only",
211
+ "status": "verified_smoke",
212
+ "processor": "bright_examples_parquet_v1",
213
+ "expected_counts": {"excluded_documents": 0, "gold": 101, "inputs": 101, "qrels": 520},
214
+ "prepared_sha256": {
215
+ "gold.jsonl": "cba0b094d5149a8ba28a9e7bd7df861f39b3a39980b26c1e527dd55c79908d7c",
216
+ "inputs.jsonl": "51460c06adf644fe25b92dca46c6360529e67c216d5c537c6dcbe7783d7512fb"
217
+ },
218
+ "allowed_hosts": ["cas-bridge.xethub.hf.co", "cdn-lfs.huggingface.co", "huggingface.co", "us.aws.cdn.hf.co"],
219
+ "provenance": {
220
+ "dataset_card": "https://huggingface.co/datasets/xlangai/BRIGHT",
221
+ "dataset_home": "https://brightbenchmark.github.io/",
222
+ "source_repository": "https://github.com/xlang-ai/BRIGHT"
223
+ },
224
+ "note": "Official Robotics queries and short-document qrels. Pair with bright.robotics.corpus; answers, reasoning and long-document qrels remain evaluator-only."
225
+ },
226
+ {
227
+ "asset_id": "bright.robotics.corpus",
228
+ "cell_id": "bright.robotics",
229
+ "source_url": "https://huggingface.co/datasets/xlangai/BRIGHT/resolve/ba248cfa7b4a12da78640f81a3dd064632e25fa6/documents/robotics-00000-of-00001.parquet",
230
+ "source_revision": "ba248cfa7b4a12da78640f81a3dd064632e25fa6",
231
+ "sha256": "2c83f286006a3b2e11a677abe88f382009c5ee79f97c1f43f6a571f3f94e6d15",
232
+ "size_bytes": 7874186,
233
+ "license_id": "CC-BY-4.0",
234
+ "redistribution": "downloader_only",
235
+ "status": "verified_smoke",
236
+ "processor": "bright_corpus_parquet_v1",
237
+ "expected_counts": {"corpus": 61961},
238
+ "prepared_sha256": {"corpus.jsonl": "58a02d6309f56a7ed345e4df31ddc325b0d8715613c36b44097ec0982301fb85"},
239
+ "allowed_hosts": ["cas-bridge.xethub.hf.co", "cdn-lfs.huggingface.co", "huggingface.co", "us.aws.cdn.hf.co"],
240
+ "provenance": {
241
+ "dataset_card": "https://huggingface.co/datasets/xlangai/BRIGHT",
242
+ "dataset_home": "https://brightbenchmark.github.io/",
243
+ "source_repository": "https://github.com/xlang-ai/BRIGHT"
244
+ },
245
+ "note": "Official compact Robotics corpus, separately pinned from its query component. Source data are downloaded on demand, never bundled."
246
+ },
247
+ {
248
+ "asset_id": "bright.stackoverflow.examples",
249
+ "cell_id": "bright.stackoverflow",
250
+ "source_url": "https://huggingface.co/datasets/xlangai/BRIGHT/resolve/1c9ff4a4090306cd8d9412b108023f1f10942864/examples/stackoverflow-00000-of-00001.parquet",
251
+ "source_revision": "1c9ff4a4090306cd8d9412b108023f1f10942864",
252
+ "sha256": "97d417ba449ef70c9c9ae2937e9df106654a2554ce1533b090cb64b998a077e1",
253
+ "size_bytes": 250458,
254
+ "license_id": "CC-BY-4.0",
255
+ "redistribution": "downloader_only",
256
+ "status": "verified_smoke",
257
+ "processor": "bright_examples_parquet_v1",
258
+ "expected_counts": {"excluded_documents": 0, "gold": 117, "inputs": 117, "qrels": 478},
259
+ "prepared_sha256": {
260
+ "gold.jsonl": "ce3d2cf8bdd6c5abf232abc61e3d3185703009c7c187b0a36bdcbdf52e8ba66b",
261
+ "inputs.jsonl": "bc746b9c20d70a27aace3c65b3eb93839db4f32c3f902dddc2e217432e19d61c"
262
+ },
263
+ "allowed_hosts": ["cas-bridge.xethub.hf.co", "cdn-lfs.huggingface.co", "huggingface.co", "us.aws.cdn.hf.co"],
264
+ "provenance": {
265
+ "dataset_card": "https://huggingface.co/datasets/xlangai/BRIGHT",
266
+ "dataset_home": "https://brightbenchmark.github.io/",
267
+ "source_repository": "https://github.com/xlang-ai/BRIGHT"
268
+ },
269
+ "note": "Official Stack Overflow queries and short-document qrels. Pair with bright.stackoverflow.corpus; answers and reasoning stay evaluator-only."
270
+ },
271
+ {
272
+ "asset_id": "bright.stackoverflow.corpus",
273
+ "cell_id": "bright.stackoverflow",
274
+ "source_url": "https://huggingface.co/datasets/xlangai/BRIGHT/resolve/ba248cfa7b4a12da78640f81a3dd064632e25fa6/documents/stackoverflow-00000-of-00001.parquet",
275
+ "source_revision": "ba248cfa7b4a12da78640f81a3dd064632e25fa6",
276
+ "sha256": "d54559692f925666c3c6b1d33a696a64ef324cf5aaeff9d6f4d11fba5cd5ac8b",
277
+ "size_bytes": 39493317,
278
+ "license_id": "CC-BY-4.0",
279
+ "redistribution": "downloader_only",
280
+ "status": "verified_smoke",
281
+ "processor": "bright_corpus_parquet_v1",
282
+ "expected_counts": {"corpus": 107081},
283
+ "prepared_sha256": {"corpus.jsonl": "3fc2ef3014573dc0c9173360c42dabc8e97384764f9bb775c11b5afbf5a7a36d"},
284
+ "allowed_hosts": ["cas-bridge.xethub.hf.co", "cdn-lfs.huggingface.co", "huggingface.co", "us.aws.cdn.hf.co"],
285
+ "provenance": {
286
+ "dataset_card": "https://huggingface.co/datasets/xlangai/BRIGHT",
287
+ "dataset_home": "https://brightbenchmark.github.io/",
288
+ "source_repository": "https://github.com/xlang-ai/BRIGHT"
289
+ },
290
+ "note": "Official compact Stack Overflow corpus; source bytes are not distributed with the toolkit."
291
+ },
292
+ {
293
+ "asset_id": "musique.answerable.dev",
294
+ "cell_id": "musique.answerable",
295
+ "source_url": "https://drive.usercontent.google.com/download?id=1tGdADlNjWFaHLeZZGShh2IRcpO6Lv24h&export=download&confirm=t",
296
+ "source_revision": "v1.0",
297
+ "sha256": "98f839bf2fd5319f5c688aed77901a6d5c30b3b9f9f691ab9a8ecafb045ee0cd",
298
+ "size_bytes": 272049578,
299
+ "license_id": "CC-BY-4.0",
300
+ "redistribution": "downloader_only",
301
+ "status": "verified_smoke",
302
+ "processor": "musique_zip_v1",
303
+ "expected_counts": {
304
+ "context_blocks": 48315,
305
+ "gold": 2417,
306
+ "inputs": 2417,
307
+ "leakage_audit_records": 6371,
308
+ "leakage_audit_sources": 5,
309
+ "support_labels": 6404
310
+ },
311
+ "prepared_sha256": {
312
+ "audit.jsonl": "4cf952385298103bf977191a31a6485a262204ff3e1f2ef6952c6fcfdcb1e398",
313
+ "gold.jsonl": "83c3dbbdc027e7385202558ec4870b636b76246b3f30ede3464d594dd1fe531e",
314
+ "inputs.jsonl": "f691cdbf8d30c1fdb9ca13e809c65ed1e648940f9f0f9d900caa95a4c26cb5cc"
315
+ },
316
+ "allowed_hosts": [
317
+ "drive.usercontent.google.com"
318
+ ],
319
+ "provenance": {
320
+ "dataset_home": "https://github.com/StonyBrookNLP/musique",
321
+ "download_index": "https://github.com/StonyBrookNLP/musique#data",
322
+ "source_repository": "https://github.com/StonyBrookNLP/musique"
323
+ },
324
+ "note": "Official MuSiQue v1.0 archive; only answerable dev is normalized. Answers, aliases, decomposition, and support labels stay outside optimizer-visible inputs."
325
+ },
326
+ {
327
+ "asset_id": "2wiki.multihop.dev",
328
+ "cell_id": "2wiki.multihop",
329
+ "source_url": "https://dl.dropboxusercontent.com/s/ms2m13252h6xubs/data_ids_april7.zip",
330
+ "source_revision": "data_ids_april7",
331
+ "sha256": "95df2bf56fdabe034e27aebc580e02264232203cf52552f9efe8a919e5529eef",
332
+ "size_bytes": 258968175,
333
+ "license_id": "Apache-2.0",
334
+ "redistribution": "downloader_only",
335
+ "status": "verified_smoke",
336
+ "processor": "two_wiki_zip_v1",
337
+ "expected_counts": {
338
+ "alias_entities": 203297,
339
+ "alias_strings": 343265,
340
+ "context_blocks": 400943,
341
+ "duplicate_context_sentences": 1496,
342
+ "empty_context_sentences": 0,
343
+ "evidence_triples": 31120,
344
+ "gold": 12576,
345
+ "inputs": 12576,
346
+ "orphan_support_labels": 70,
347
+ "support_labels": 30687
348
+ },
349
+ "prepared_sha256": {
350
+ "audit.jsonl": "7a135219336c4fb828009e911e099667b2171de35e2c25e61feb3688625708e5",
351
+ "gold.jsonl": "da09635414ea70fd4d37624097af3d58bf56603bd16282e6747c95d6c38203d7",
352
+ "inputs.jsonl": "1f4cb07fc18e0994c435db53b49d4e3d49098ed1f53913edab4884198843ff45"
353
+ },
354
+ "allowed_hosts": [
355
+ "dl.dropboxusercontent.com"
356
+ ],
357
+ "provenance": {
358
+ "dataset_home": "https://github.com/Alab-NII/2wikimultihop",
359
+ "official_download": "https://www.dropbox.com/s/ms2m13252h6xubs/data_ids_april7.zip",
360
+ "source_repository": "https://github.com/Alab-NII/2wikimultihop"
361
+ },
362
+ "note": "Official corrected data_ids_april7 archive with v1.1 answer/evidence IDs and alias table. Only dev is normalized; train/test remain source-verified but are not emitted."
363
+ },
364
+ {
365
+ "asset_id": "hotpotqa.distractor.validation",
366
+ "cell_id": "hotpotqa.distractor",
367
+ "source_url": "https://huggingface.co/datasets/hotpotqa/hotpot_qa/resolve/1908d6afbbead072334abe2965f91bd2709910ab/distractor/validation-00000-of-00001.parquet",
368
+ "source_revision": "1908d6afbbead072334abe2965f91bd2709910ab",
369
+ "sha256": "c20b638ca82b21d04fe12e14ff417ad05153d4d215a65de54497fca4e972f7c6",
370
+ "size_bytes": 27452575,
371
+ "license_id": "CC-BY-SA-4.0",
372
+ "redistribution": "downloader_only",
373
+ "status": "verified_smoke",
374
+ "processor": "hotpotqa_parquet_v1",
375
+ "expected_counts": {
376
+ "context_blocks": 306438,
377
+ "duplicate_context_sentences": 0,
378
+ "empty_context_sentences": 49,
379
+ "gold": 7405,
380
+ "inputs": 7405,
381
+ "orphan_support_labels": 1,
382
+ "support_labels": 18005
383
+ },
384
+ "prepared_sha256": {
385
+ "gold.jsonl": "d53aa1257e243be256dee6751f7a4683113f52f619d76598c28bca3e9d2ff655",
386
+ "inputs.jsonl": "92c61fbf5661a9227b54c5e3594f932dd87ca39cbbf05602099cdb32ce2ebd72"
387
+ },
388
+ "allowed_hosts": [
389
+ "cas-bridge.xethub.hf.co",
390
+ "cdn-lfs.huggingface.co",
391
+ "huggingface.co",
392
+ "us.aws.cdn.hf.co"
393
+ ],
394
+ "provenance": {
395
+ "dataset_card": "https://huggingface.co/datasets/hotpotqa/hotpot_qa",
396
+ "dataset_home": "https://hotpotqa.github.io/",
397
+ "source_repository": "https://github.com/hotpotqa/hotpot"
398
+ },
399
+ "note": "Official Hugging Face organization mirror of the distractor validation split. One impossible upstream supporting-fact index is preserved and explicitly audited."
400
+ },
401
+ {
402
+ "asset_id": "qasper.dev.official",
403
+ "cell_id": "qasper.v0_3_dev",
404
+ "source_url": "https://qasper-dataset.s3.us-west-2.amazonaws.com/qasper-train-dev-v0.3.tgz",
405
+ "source_revision": "v0.3",
406
+ "sha256": "a28fdf966db827bcee3d873107d6b6669864fb7ca8fbf73a192f5e39191bdb5a",
407
+ "size_bytes": 10835856,
408
+ "license_id": "CC-BY-4.0",
409
+ "redistribution": "downloader_only",
410
+ "status": "verified_smoke",
411
+ "processor": "qasper_tgz_v1",
412
+ "expected_counts": {
413
+ "annotations": 1764,
414
+ "context_blocks": 66990,
415
+ "evidence_labels": 2808,
416
+ "gold": 1005,
417
+ "inputs": 1005,
418
+ "papers": 281,
419
+ "unique_paper_context_blocks": 18990
420
+ },
421
+ "prepared_sha256": {
422
+ "gold.jsonl": "b73b997187e5f05941cde19194e2338d6e7695d3270e133136da77ac9dcb8d9a",
423
+ "inputs.jsonl": "a0475bc93b9551fdb57f02213c88b06a373126cca545a5e50da53df6542accdf"
424
+ },
425
+ "allowed_hosts": [
426
+ "qasper-dataset.s3.us-west-2.amazonaws.com"
427
+ ],
428
+ "provenance": {
429
+ "dataset_card": "https://huggingface.co/datasets/allenai/qasper",
430
+ "dataset_home": "https://allenai.org/data/qasper",
431
+ "source_repository": "https://github.com/allenai/qasper-led-baseline"
432
+ },
433
+ "note": "Official QASPER v0.3 train/dev archive; only dev is normalized. The locked setting is caption-augmented full evidence (including FLOAT SELECTED captions), distinct from the LED text-only baseline."
434
+ },
435
+ {
436
+ "asset_id": "tau3.banking_knowledge.v1_0_1",
437
+ "cell_id": "tau3.banking_knowledge",
438
+ "source_url": "https://github.com/sierra-research/tau2-bench/archive/cf71a8070269883e38a365ffa85f78f46844c1f4.tar.gz",
439
+ "source_revision": "cf71a8070269883e38a365ffa85f78f46844c1f4",
440
+ "sha256": "c8e8dafdcbf3671f5129705b420a07dd9516f1f962ea6484809532ace1d7c708",
441
+ "size_bytes": 93248469,
442
+ "license_id": "MIT",
443
+ "redistribution": "downloader_only",
444
+ "status": "verified_smoke",
445
+ "processor": "tau3_banking_tar_v1",
446
+ "expected_counts": {
447
+ "audit": 1,
448
+ "corpus": 698,
449
+ "documents": 698,
450
+ "gold": 97,
451
+ "gold_actions": 955,
452
+ "inputs": 97,
453
+ "legacy_combined_task_mismatches": 13,
454
+ "required_document_labels": 959,
455
+ "required_document_unique": 218,
456
+ "tasks": 97,
457
+ "tasks_with_initial_state": 75,
458
+ "tasks_with_initialization_data": 53,
459
+ "tasks_with_message_history": 0,
460
+ "user_simulator_scripts": 97
461
+ },
462
+ "prepared_sha256": {
463
+ "audit.jsonl": "6aad364f6ab027bd1eac483a351b26b6a2348f1b61925dd4a64fac5b979826f9",
464
+ "corpus.jsonl": "c169e567d4f24b68345e128010c171caa990473300fe6e265db0b7e4f2d5582d",
465
+ "gold.jsonl": "1df4287a451701a6b837cd04126ce67a54417151c0e09904a618799e4363ad4a",
466
+ "inputs.jsonl": "7653ca07e93098ac9cba96fe3bcd565560fa51f793e51c5fc9d162f991bc1b35"
467
+ },
468
+ "allowed_hosts": [
469
+ "codeload.github.com",
470
+ "github.com"
471
+ ],
472
+ "provenance": {
473
+ "dataset_home": "https://taubench.com/",
474
+ "paper": "https://openreview.net/forum?id=XHZK5abtw2",
475
+ "source_repository": "https://github.com/sierra-research/tau2-bench",
476
+ "source_tree": "https://github.com/sierra-research/tau2-bench/tree/cf71a8070269883e38a365ffa85f78f46844c1f4/data/tau2/domains/banking_knowledge"
477
+ },
478
+ "note": "Official tau2 1.0.1 Banking Knowledge source. Individual task files are authoritative; static tasks contain hidden simulator/evaluator state and therefore require explicit replayable visible checkpoints before intervention. Source bytes and prepared rows are not packaged."
479
+ },
480
+ {
481
+ "asset_id": "clariq.dev.official",
482
+ "cell_id": "clariq.main",
483
+ "source_url": "https://raw.githubusercontent.com/aliannejadi/ClariQ/46885a544581a0af8aff0681d29e4971807e2912/data/dev.tsv",
484
+ "source_revision": "46885a544581a0af8aff0681d29e4971807e2912",
485
+ "sha256": "68d2a5f87eab73721979b5f45f64099a9b2f080db1d0ce4b979d9daa4249906e",
486
+ "size_bytes": 585998,
487
+ "license_id": "NOASSERTION",
488
+ "redistribution": "not_redistributed",
489
+ "status": "license_review_required",
490
+ "processor": "clariq_tsv_v1",
491
+ "expected_counts": {
492
+ "gold": 2313,
493
+ "inputs": 2313,
494
+ "topics": 50
495
+ },
496
+ "prepared_sha256": {
497
+ "gold.jsonl": "2ff6a5d45c87cc3f4cf55e1d78628a97c9969593e2bc8d0e101d23ea45d41330",
498
+ "inputs.jsonl": "bbf22cb67f1c21866074e85737d365fe38fcc4b34e1009b6aa8465db7d256670"
499
+ },
500
+ "allowed_hosts": [
501
+ "raw.githubusercontent.com"
502
+ ],
503
+ "provenance": {
504
+ "dataset_home": "https://github.com/aliannejadi/ClariQ",
505
+ "source_repository": "https://github.com/aliannejadi/ClariQ",
506
+ "source_tree": "https://github.com/aliannejadi/ClariQ/tree/46885a544581a0af8aff0681d29e4971807e2912/data"
507
+ },
508
+ "note": "Public official dev file, but the upstream repository publishes no dataset license; explicit local-use acceptance is required and no rows are redistributed."
509
+ }
510
+ ]
511
+ }