synth-ai 0.2.9.dev0__py3-none-any.whl → 0.2.23.dev3__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 (890) hide show
  1. examples/README.md +1 -0
  2. examples/__init__.py +16 -0
  3. examples/analyze_semantic_words.sh +17 -0
  4. examples/baseline/banking77_baseline.py +243 -0
  5. examples/baseline/banking77_pipeline_baseline.py +294 -0
  6. examples/baseline/crafter_baseline.py +407 -0
  7. examples/baseline/pokemon_red_baseline.py +326 -0
  8. examples/baseline/simple_baseline.py +56 -0
  9. examples/baseline/warming_up_to_rl_baseline.py +239 -0
  10. examples/blog_posts/gepa/README.md +355 -0
  11. examples/blog_posts/gepa/configs/banking77_gepa_local.toml +95 -0
  12. examples/blog_posts/gepa/configs/banking77_gepa_test.toml +80 -0
  13. examples/blog_posts/gepa/configs/banking77_mipro_local.toml +50 -0
  14. examples/blog_posts/gepa/configs/banking77_pipeline_gepa_local.toml +101 -0
  15. examples/blog_posts/gepa/configs/banking77_pipeline_gepa_test.toml +96 -0
  16. examples/blog_posts/gepa/configs/hotpotqa_gepa_local.toml +57 -0
  17. examples/blog_posts/gepa/configs/hotpotqa_gepa_qwen.toml +35 -0
  18. examples/blog_posts/gepa/configs/hotpotqa_mipro_local.toml +51 -0
  19. examples/blog_posts/gepa/configs/hover_gepa_local.toml +57 -0
  20. examples/blog_posts/gepa/configs/hover_gepa_qwen.toml +35 -0
  21. examples/blog_posts/gepa/configs/hover_mipro_local.toml +51 -0
  22. examples/blog_posts/gepa/configs/ifbench_gepa_local.toml +57 -0
  23. examples/blog_posts/gepa/configs/ifbench_gepa_qwen.toml +35 -0
  24. examples/blog_posts/gepa/configs/ifbench_mipro_local.toml +51 -0
  25. examples/blog_posts/gepa/configs/pupa_gepa_local.toml +58 -0
  26. examples/blog_posts/gepa/configs/pupa_mipro_local.toml +52 -0
  27. examples/blog_posts/gepa/deploy_banking77_task_app.sh +54 -0
  28. examples/blog_posts/gepa/gepa_baseline.py +204 -0
  29. examples/blog_posts/gepa/query_prompts_example.py +97 -0
  30. examples/blog_posts/gepa/run_gepa_banking77.sh +112 -0
  31. examples/blog_posts/gepa/run_gepa_banking77_pipeline.sh +163 -0
  32. examples/blog_posts/gepa/task_apps.py +105 -0
  33. examples/blog_posts/gepa/test_gepa_local.sh +67 -0
  34. examples/blog_posts/gepa/verify_banking77_setup.sh +123 -0
  35. examples/blog_posts/mipro/README.md +415 -0
  36. examples/blog_posts/mipro/configs/banking77_mipro_local.toml +91 -0
  37. examples/blog_posts/mipro/configs/banking77_mipro_test.toml +87 -0
  38. examples/blog_posts/mipro/configs/banking77_pipeline_mipro_gemini_flash_lite_local.toml +98 -0
  39. examples/blog_posts/mipro/configs/banking77_pipeline_mipro_gpt41mini_local.toml +96 -0
  40. examples/blog_posts/mipro/configs/banking77_pipeline_mipro_local.toml +94 -0
  41. examples/blog_posts/mipro/configs/banking77_pipeline_mipro_test.toml +170 -0
  42. examples/blog_posts/mipro/deploy_banking77_pipeline_task_app.sh +59 -0
  43. examples/blog_posts/mipro/deploy_banking77_task_app.sh +41 -0
  44. examples/blog_posts/mipro/multi_step.md +79 -0
  45. examples/blog_posts/mipro/run_mipro_banking77.sh +191 -0
  46. examples/blog_posts/mipro/run_mipro_banking77_pipeline.sh +171 -0
  47. examples/blog_posts/mipro/run_mipro_banking77_pipeline_gemini_flash_lite.sh +177 -0
  48. examples/blog_posts/mipro/run_mipro_banking77_pipeline_gpt41mini.sh +173 -0
  49. examples/blog_posts/mipro/verify_banking77_setup.sh +117 -0
  50. examples/blog_posts/pokemon_vl/README.md +98 -0
  51. examples/blog_posts/pokemon_vl/configs/eval_gpt5nano.toml +26 -0
  52. examples/blog_posts/pokemon_vl/configs/eval_qwen3_vl.toml +27 -0
  53. examples/blog_posts/pokemon_vl/configs/eval_rl_final.toml +24 -0
  54. examples/blog_posts/pokemon_vl/configs/filter_high_reward.toml +10 -0
  55. examples/blog_posts/pokemon_vl/configs/train_rl_from_sft.toml +43 -0
  56. examples/blog_posts/pokemon_vl/configs/train_sft_qwen4b_vl.toml +40 -0
  57. examples/blog_posts/pokemon_vl/extract_images.py +239 -0
  58. examples/blog_posts/pokemon_vl/pokemon_vl_baseline.py +326 -0
  59. examples/blog_posts/pokemon_vl/run_eval_extract_images.py +209 -0
  60. examples/blog_posts/pokemon_vl/run_qwen_eval_extract_images.py +212 -0
  61. examples/blog_posts/pokemon_vl/text_box_analysis.md +106 -0
  62. examples/blog_posts/warming_up_to_rl/ARCHITECTURE.md +195 -0
  63. examples/blog_posts/warming_up_to_rl/FINAL_TEST_RESULTS.md +127 -0
  64. examples/blog_posts/warming_up_to_rl/INFERENCE_SUCCESS.md +132 -0
  65. examples/blog_posts/warming_up_to_rl/README.md +158 -0
  66. examples/blog_posts/warming_up_to_rl/SMOKE_TESTING.md +164 -0
  67. examples/blog_posts/warming_up_to_rl/SMOKE_TEST_COMPLETE.md +253 -0
  68. examples/blog_posts/warming_up_to_rl/configs/eval_baseline_qwen32b_10x20.toml +25 -0
  69. examples/blog_posts/warming_up_to_rl/configs/eval_ft_qwen4b.toml +25 -0
  70. examples/blog_posts/warming_up_to_rl/configs/eval_ft_qwen4b_10x20.toml +26 -0
  71. examples/blog_posts/warming_up_to_rl/configs/eval_groq_qwen32b.toml +25 -0
  72. examples/blog_posts/warming_up_to_rl/configs/eval_openai_gpt_oss_120b.toml +29 -0
  73. examples/blog_posts/warming_up_to_rl/configs/filter_high_reward_dataset.toml +10 -0
  74. examples/blog_posts/warming_up_to_rl/configs/smoke_test.toml +75 -0
  75. examples/blog_posts/warming_up_to_rl/configs/train_rl_from_sft.toml +91 -0
  76. examples/blog_posts/warming_up_to_rl/configs/train_sft_qwen4b.toml +40 -0
  77. examples/blog_posts/warming_up_to_rl/warming_up_to_rl_baseline.py +187 -0
  78. examples/crafter_debug_render.py +186 -0
  79. examples/dev/qwen3_32b_qlora_4xh100.toml +45 -0
  80. examples/gepa/banking77_pipeline_gepa.toml +96 -0
  81. examples/gepa/multi_stage_gepa_example.toml +84 -0
  82. examples/gepa/run_gepa_banking77_pipeline.sh +157 -0
  83. examples/multi_step/SFT_README.md +147 -0
  84. examples/multi_step/configs/README_verilog_rl.md +77 -0
  85. examples/multi_step/configs/VERILOG_REWARDS.md +103 -0
  86. examples/multi_step/configs/VERILOG_RL_CHECKLIST.md +196 -0
  87. examples/multi_step/configs/crafter_eval_synth_qwen4b.toml +35 -0
  88. examples/multi_step/configs/crafter_eval_text_only_groq_qwen32b.toml +36 -0
  89. examples/multi_step/configs/crafter_rl_outcome.toml +75 -0
  90. examples/multi_step/configs/crafter_rl_stepwise_hosted_judge.toml +145 -0
  91. examples/multi_step/configs/crafter_rl_stepwise_shaped.toml +84 -0
  92. examples/multi_step/configs/crafter_rl_stepwise_simple.toml +79 -0
  93. examples/multi_step/configs/crafter_rl_stepwise_simple_NEW_FORMAT.toml +105 -0
  94. examples/multi_step/configs/crafter_sft_qwen30b_lora.toml +62 -0
  95. examples/multi_step/configs/crafter_synth_backend.md +40 -0
  96. examples/multi_step/configs/verilog_eval_groq_qwen32b.toml +31 -0
  97. examples/multi_step/configs/verilog_eval_synth_qwen8b.toml +33 -0
  98. examples/multi_step/configs/verilog_rl_lora.toml +147 -0
  99. examples/multi_step/convert_traces_to_sft.py +84 -0
  100. examples/multi_step/crafter_rl_lora.md +70 -0
  101. examples/multi_step/judges/crafter_backend_judge.py +220 -0
  102. examples/multi_step/judges/verilog_backend_judge.py +234 -0
  103. examples/multi_step/readme.md +48 -0
  104. examples/multi_step/run_sft_qwen30b.sh +45 -0
  105. examples/multi_step/sse_metrics_streaming_notes.md +357 -0
  106. examples/multi_step/task_app_config_notes.md +494 -0
  107. examples/multi_step/verilog_rl_lora.md +218 -0
  108. examples/qwen_coder/README.md +102 -0
  109. examples/qwen_coder/_shared.py +113 -0
  110. examples/qwen_coder/configs/coder_lora_30b.toml +60 -0
  111. examples/qwen_coder/configs/coder_lora_4b.toml +61 -0
  112. examples/qwen_coder/configs/coder_lora_small.toml +57 -0
  113. examples/qwen_coder/generate_dataset.py +98 -0
  114. examples/qwen_coder/infer_ft_smoke.py +65 -0
  115. examples/qwen_coder/infer_prod_proxy.py +73 -0
  116. examples/qwen_coder/infer_via_synth.py +87 -0
  117. examples/qwen_coder/scripts/infer_coder.sh +19 -0
  118. examples/qwen_coder/scripts/train_coder_30b.sh +22 -0
  119. examples/qwen_coder/sft_full_17b.py +103 -0
  120. examples/qwen_coder/sft_lora_30b.py +110 -0
  121. examples/qwen_coder/subset_jsonl.py +39 -0
  122. examples/qwen_coder/todos.md +38 -0
  123. examples/qwen_coder/validate_jsonl.py +60 -0
  124. examples/qwen_vl/BUGS_AND_FIXES.md +232 -0
  125. examples/qwen_vl/IMAGE_VALIDATION_COMPLETE.md +271 -0
  126. examples/qwen_vl/IMAGE_VALIDATION_SUMMARY.md +260 -0
  127. examples/qwen_vl/INFERENCE_SFT_TESTS.md +412 -0
  128. examples/qwen_vl/NEXT_STEPS_2B.md +325 -0
  129. examples/qwen_vl/QUICKSTART.md +327 -0
  130. examples/qwen_vl/QUICKSTART_RL_VISION.md +110 -0
  131. examples/qwen_vl/README.md +152 -0
  132. examples/qwen_vl/RL_VISION_COMPLETE.md +475 -0
  133. examples/qwen_vl/RL_VISION_TESTING.md +333 -0
  134. examples/qwen_vl/SDK_VISION_INTEGRATION.md +328 -0
  135. examples/qwen_vl/SETUP_COMPLETE.md +274 -0
  136. examples/qwen_vl/VISION_TESTS_COMPLETE.md +489 -0
  137. examples/qwen_vl/VLM_PIPELINE_COMPLETE.md +242 -0
  138. examples/qwen_vl/__init__.py +2 -0
  139. examples/qwen_vl/collect_data_via_cli.md +415 -0
  140. examples/qwen_vl/collect_vision_traces.py +368 -0
  141. examples/qwen_vl/configs/crafter_rl_vision_qwen3vl4b.toml +110 -0
  142. examples/qwen_vl/configs/crafter_vlm_sft_example.toml +59 -0
  143. examples/qwen_vl/configs/eval_gpt4o_mini_vision.toml +26 -0
  144. examples/qwen_vl/configs/eval_gpt4o_vision_proper.toml +29 -0
  145. examples/qwen_vl/configs/eval_gpt5nano_vision.toml +26 -0
  146. examples/qwen_vl/configs/eval_qwen3vl_vision.toml +26 -0
  147. examples/qwen_vl/configs/filter_qwen3vl_sft.toml +49 -0
  148. examples/qwen_vl/configs/filter_vision_sft.toml +52 -0
  149. examples/qwen_vl/configs/filter_vision_test.toml +8 -0
  150. examples/qwen_vl/configs/sft_qwen3_vl_2b_test.toml +54 -0
  151. examples/qwen_vl/crafter_gpt5nano_agent.py +308 -0
  152. examples/qwen_vl/crafter_qwen_vl_agent.py +300 -0
  153. examples/qwen_vl/run_vision_comparison.sh +61 -0
  154. examples/qwen_vl/run_vision_sft_pipeline.sh +175 -0
  155. examples/qwen_vl/test_image_validation.py +201 -0
  156. examples/qwen_vl/test_sft_vision_data.py +110 -0
  157. examples/rl/README.md +169 -0
  158. examples/rl/configs/eval_base_qwen.toml +17 -0
  159. examples/rl/configs/eval_rl_qwen.toml +13 -0
  160. examples/rl/configs/rl_from_base_qwen.toml +62 -0
  161. examples/rl/configs/rl_from_base_qwen17.toml +80 -0
  162. examples/rl/configs/rl_from_ft_qwen.toml +37 -0
  163. examples/rl/download_dataset.py +80 -0
  164. examples/rl/run_eval.py +436 -0
  165. examples/rl/run_rl_and_save.py +111 -0
  166. examples/rl/task_app/README.md +21 -0
  167. {synth_ai/task/apps → examples/rl/task_app}/math_single_step.py +188 -50
  168. examples/rl/task_app/math_task_app.py +111 -0
  169. examples/run_crafter_demo.sh +10 -0
  170. examples/sdk_prompt_learning_example.py +55 -0
  171. examples/sft/README.md +139 -0
  172. examples/sft/configs/crafter_fft_qwen0p6b.toml +49 -0
  173. examples/sft/configs/crafter_lora_qwen0p6b.toml +49 -0
  174. examples/sft/evaluate.py +117 -0
  175. examples/sft/export_dataset.py +120 -0
  176. examples/sft/generate_traces.py +164 -0
  177. examples/swe/__init__.py +12 -0
  178. examples/swe/task_app/README.md +135 -0
  179. examples/swe/task_app/__init__.py +2 -0
  180. examples/swe/task_app/grpo_swe_mini.py +604 -0
  181. examples/swe/task_app/grpo_swe_mini_task_app.py +124 -0
  182. examples/swe/task_app/hosted/README.md +173 -0
  183. examples/swe/task_app/hosted/__init__.py +5 -0
  184. examples/swe/task_app/hosted/branching.py +143 -0
  185. examples/swe/task_app/hosted/environment_routes.py +1289 -0
  186. examples/swe/task_app/hosted/envs/__init__.py +1 -0
  187. examples/swe/task_app/hosted/envs/crafter/__init__.py +6 -0
  188. examples/swe/task_app/hosted/envs/crafter/app.py +1 -0
  189. examples/swe/task_app/hosted/envs/crafter/environment.py +522 -0
  190. examples/swe/task_app/hosted/envs/crafter/policy.py +478 -0
  191. examples/swe/task_app/hosted/envs/crafter/react_agent.py +108 -0
  192. examples/swe/task_app/hosted/envs/crafter/shared.py +305 -0
  193. examples/swe/task_app/hosted/envs/crafter/tools.py +47 -0
  194. examples/swe/task_app/hosted/envs/mini_swe/__init__.py +8 -0
  195. examples/swe/task_app/hosted/envs/mini_swe/environment.py +1191 -0
  196. examples/swe/task_app/hosted/envs/mini_swe/policy.py +355 -0
  197. examples/swe/task_app/hosted/envs/mini_swe/shared.py +83 -0
  198. examples/swe/task_app/hosted/envs/mini_swe/tools.py +96 -0
  199. examples/swe/task_app/hosted/hosted_app.py +204 -0
  200. examples/swe/task_app/hosted/inference/__init__.py +5 -0
  201. examples/swe/task_app/hosted/inference/openai_client.py +584 -0
  202. examples/swe/task_app/hosted/main.py +100 -0
  203. examples/swe/task_app/hosted/policy_routes.py +1094 -0
  204. examples/swe/task_app/hosted/registry.py +195 -0
  205. examples/swe/task_app/hosted/rollout.py +1905 -0
  206. examples/swe/task_app/hosted/storage/__init__.py +5 -0
  207. examples/swe/task_app/hosted/storage/volume.py +211 -0
  208. examples/swe/task_app/hosted/test_agents.py +161 -0
  209. examples/swe/task_app/hosted/test_service.py +136 -0
  210. examples/swe/task_app/hosted/utils.py +62 -0
  211. examples/swe/task_app/morph_backend.py +178 -0
  212. examples/task_apps/IMAGE_ONLY_EVAL_QUICKSTART.md +258 -0
  213. examples/task_apps/TESTING.md +275 -0
  214. examples/task_apps/banking77/__init__.py +6 -0
  215. examples/task_apps/banking77/banking77_task_app.py +912 -0
  216. examples/task_apps/banking77/deploy_wrapper.py +46 -0
  217. examples/task_apps/banking77_pipeline/__init__.py +6 -0
  218. examples/task_apps/banking77_pipeline/banking77_pipeline_task_app.py +489 -0
  219. examples/task_apps/banking77_pipeline/deploy_wrapper.py +50 -0
  220. examples/task_apps/crafter/CREATE_SFT_DATASET.md +286 -0
  221. examples/task_apps/crafter/EVAL_IMAGE_ONLY_RESULTS.md +152 -0
  222. examples/task_apps/crafter/FILTER_COMMAND_STATUS.md +187 -0
  223. examples/task_apps/crafter/FILTER_COMMAND_SUCCESS.md +281 -0
  224. examples/task_apps/crafter/QUERY_EXAMPLES.md +203 -0
  225. examples/task_apps/crafter/README_IMAGE_ONLY_EVAL.md +316 -0
  226. examples/task_apps/crafter/eval_image_only_gpt4o.toml +28 -0
  227. examples/task_apps/crafter/eval_text_only_groq_llama.toml +36 -0
  228. examples/task_apps/crafter/filter_sft_dataset.toml +16 -0
  229. examples/task_apps/crafter/task_app/README.md +42 -0
  230. examples/task_apps/crafter/task_app/__init__.py +5 -0
  231. examples/task_apps/crafter/task_app/grpo_crafter.py +1055 -0
  232. examples/task_apps/crafter/task_app/grpo_crafter_task_app.py +146 -0
  233. examples/task_apps/crafter/task_app/synth_envs_hosted/README.md +173 -0
  234. examples/task_apps/crafter/task_app/synth_envs_hosted/__init__.py +5 -0
  235. examples/task_apps/crafter/task_app/synth_envs_hosted/branching.py +143 -0
  236. examples/task_apps/crafter/task_app/synth_envs_hosted/environment_routes.py +1226 -0
  237. examples/task_apps/crafter/task_app/synth_envs_hosted/envs/__init__.py +1 -0
  238. examples/task_apps/crafter/task_app/synth_envs_hosted/envs/crafter/__init__.py +6 -0
  239. examples/task_apps/crafter/task_app/synth_envs_hosted/envs/crafter/app.py +1 -0
  240. examples/task_apps/crafter/task_app/synth_envs_hosted/envs/crafter/environment.py +532 -0
  241. examples/task_apps/crafter/task_app/synth_envs_hosted/envs/crafter/policy.py +583 -0
  242. examples/task_apps/crafter/task_app/synth_envs_hosted/envs/crafter/react_agent.py +122 -0
  243. examples/task_apps/crafter/task_app/synth_envs_hosted/envs/crafter/shared.py +305 -0
  244. examples/task_apps/crafter/task_app/synth_envs_hosted/envs/crafter/tools.py +47 -0
  245. examples/task_apps/crafter/task_app/synth_envs_hosted/hosted_app.py +253 -0
  246. examples/task_apps/crafter/task_app/synth_envs_hosted/inference/__init__.py +5 -0
  247. examples/task_apps/crafter/task_app/synth_envs_hosted/inference/openai_client.py +999 -0
  248. examples/task_apps/crafter/task_app/synth_envs_hosted/main.py +100 -0
  249. examples/task_apps/crafter/task_app/synth_envs_hosted/policy_routes.py +1252 -0
  250. examples/task_apps/crafter/task_app/synth_envs_hosted/registry.py +195 -0
  251. examples/task_apps/crafter/task_app/synth_envs_hosted/rollout.py +2233 -0
  252. examples/task_apps/crafter/task_app/synth_envs_hosted/storage/__init__.py +5 -0
  253. examples/task_apps/crafter/task_app/synth_envs_hosted/storage/volume.py +211 -0
  254. examples/task_apps/crafter/task_app/synth_envs_hosted/test_agents.py +161 -0
  255. examples/task_apps/crafter/task_app/synth_envs_hosted/test_service.py +136 -0
  256. examples/task_apps/crafter/task_app/synth_envs_hosted/utils.py +411 -0
  257. examples/task_apps/dev/pokemon_emerald/__init__.py +2 -0
  258. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/README.md +811 -0
  259. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/agent/__init__.py +120 -0
  260. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/agent/action.py +160 -0
  261. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/agent/memory.py +155 -0
  262. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/agent/perception.py +69 -0
  263. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/agent/planning.py +96 -0
  264. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/agent/simple.py +1502 -0
  265. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/agent/system_prompt.py +4 -0
  266. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/grab_map.py +68 -0
  267. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/manual.py +216 -0
  268. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/pokemon_env/__init__.py +35 -0
  269. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/pokemon_env/emerald_utils.py +631 -0
  270. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/pokemon_env/emulator.py +1544 -0
  271. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/pokemon_env/enums.py +1428 -0
  272. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/pokemon_env/memory_reader.py +4848 -0
  273. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/pokemon_env/types.py +41 -0
  274. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/pokemon_env/utils.py +298 -0
  275. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/pyproject.toml +95 -0
  276. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/run.py +204 -0
  277. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/server/app.py +2152 -0
  278. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/server/client.py +429 -0
  279. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/server/frame_server.py +155 -0
  280. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/tests/README.md +78 -0
  281. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/tests/run_tests.py +122 -0
  282. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/tests/test_agent_direct.py +76 -0
  283. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/tests/test_agent_prompts.py +413 -0
  284. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/tests/test_battle_state_formatting.py +204 -0
  285. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/tests/test_dialogue_detection.py +133 -0
  286. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/tests/test_dialogue_detection_comprehensive.py +229 -0
  287. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/tests/test_direct_agent_emulator.py +300 -0
  288. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/tests/test_fps_adjustment_pytest.py +205 -0
  289. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/tests/test_house_to_outside_direct.py +200 -0
  290. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/tests/test_house_to_outside_transition.py +284 -0
  291. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/tests/test_map_ground_truth_comparison.py +468 -0
  292. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/tests/test_memory_map.py +575 -0
  293. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/tests/test_server_map_validation.py +311 -0
  294. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/tests/test_torchic_state.py +259 -0
  295. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/utils/anticheat.py +372 -0
  296. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/utils/checkpoint.py +296 -0
  297. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/utils/error_handler.py +275 -0
  298. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/utils/get_local_ip.py +22 -0
  299. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/utils/helpers.py +44 -0
  300. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/utils/llm_logger.py +514 -0
  301. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/utils/map_formatter.py +415 -0
  302. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/utils/map_stitcher.py +1763 -0
  303. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/utils/map_stitcher_singleton.py +33 -0
  304. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/utils/map_trimmer.py +106 -0
  305. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/utils/map_visualizer.py +334 -0
  306. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/utils/ocr_dialogue.py +1020 -0
  307. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/utils/recording.py +188 -0
  308. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/utils/state_formatter.py +1481 -0
  309. examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/utils/vlm.py +862 -0
  310. examples/task_apps/dev/pokemon_emerald/modal_app.py +114 -0
  311. examples/task_apps/dev/pokemon_emerald/task_app/README.md +81 -0
  312. examples/task_apps/dev/pokemon_emerald/task_app/__init__.py +6 -0
  313. examples/task_apps/dev/pokemon_emerald/task_app/pokemon_emerald.py +685 -0
  314. examples/task_apps/enron/__init__.py +2 -0
  315. examples/task_apps/enron/eval_groq_qwen32.toml +16 -0
  316. examples/task_apps/enron/filter_sft.toml +5 -0
  317. examples/task_apps/enron/task_app/README.md +14 -0
  318. examples/task_apps/enron/task_app/__init__.py +1 -0
  319. examples/task_apps/enron/task_app/grpo_enron.py +906 -0
  320. examples/task_apps/enron/task_app/grpo_enron_task_app.py +146 -0
  321. examples/task_apps/enron/tests/__init__.py +4 -0
  322. examples/task_apps/enron/tests/conftest.py +115 -0
  323. examples/task_apps/enron/tests/integration/__init__.py +4 -0
  324. examples/task_apps/enron/tests/integration/test_enron_eval.py +179 -0
  325. examples/task_apps/enron/tests/integration/test_enron_rollout.py +135 -0
  326. examples/task_apps/enron/tests/unit/__init__.py +4 -0
  327. examples/task_apps/enron/tests/unit/test_enron_environment.py +126 -0
  328. examples/task_apps/gepa_benchmarks/__init__.py +7 -0
  329. examples/task_apps/gepa_benchmarks/common.py +260 -0
  330. examples/task_apps/gepa_benchmarks/hotpotqa_task_app.py +507 -0
  331. examples/task_apps/gepa_benchmarks/hover_task_app.py +436 -0
  332. examples/task_apps/gepa_benchmarks/ifbench_task_app.py +563 -0
  333. examples/task_apps/gepa_benchmarks/pupa_task_app.py +460 -0
  334. examples/task_apps/math/README.md +21 -0
  335. examples/task_apps/math/math_single_step.py +1000 -0
  336. examples/task_apps/math/math_task_app.py +115 -0
  337. examples/task_apps/pokemon_battle/__init__.py +2 -0
  338. examples/task_apps/pokemon_battle/modal_app.py +104 -0
  339. examples/task_apps/pokemon_battle/task_app/README.md +68 -0
  340. examples/task_apps/pokemon_battle/task_app/__init__.py +6 -0
  341. examples/task_apps/pokemon_battle/task_app/pokemon_showdown.py +932 -0
  342. examples/task_apps/pokemon_red/EVAL_IMAGE_ONLY_COMPLETE.md +283 -0
  343. examples/task_apps/pokemon_red/EVAL_IMAGE_ONLY_STATUS.md +155 -0
  344. examples/task_apps/pokemon_red/README.md +356 -0
  345. examples/task_apps/pokemon_red/README_IMAGE_ONLY_EVAL.md +428 -0
  346. examples/task_apps/pokemon_red/__init__.py +3 -0
  347. examples/task_apps/pokemon_red/eval_image_only_gpt4o.toml +30 -0
  348. examples/task_apps/pokemon_red/eval_pokemon_red_policy.py +224 -0
  349. examples/task_apps/pokemon_red/pallet_town_rl_config.toml +75 -0
  350. examples/task_apps/pokemon_red/task_app.py +1048 -0
  351. examples/task_apps/pokemon_red/test_pallet_town_rewards.py +193 -0
  352. examples/task_apps/sokoban/README.md +306 -0
  353. examples/task_apps/sokoban/__init__.py +3 -0
  354. examples/task_apps/sokoban/eval_groq_qwen32.toml +16 -0
  355. examples/task_apps/sokoban/eval_openai_gpt5.toml +16 -0
  356. examples/task_apps/sokoban/filter_sft.toml +5 -0
  357. examples/task_apps/sokoban/task_app.py +1058 -0
  358. examples/task_apps/sokoban/tests/__init__.py +4 -0
  359. examples/task_apps/sokoban/tests/conftest.py +113 -0
  360. examples/task_apps/sokoban/tests/integration/__init__.py +4 -0
  361. examples/task_apps/sokoban/tests/integration/test_sokoban_eval.py +57 -0
  362. examples/task_apps/sokoban/tests/integration/test_sokoban_rollout.py +198 -0
  363. examples/task_apps/sokoban/tests/unit/__init__.py +4 -0
  364. examples/task_apps/sokoban/tests/unit/test_sokoban_environment.py +114 -0
  365. examples/task_apps/verilog/__init__.py +1 -0
  366. examples/task_apps/verilog/eval_groq_qwen32b.toml +22 -0
  367. examples/task_apps/verilog/filter_sft.toml +5 -0
  368. examples/task_apps/verilog/task_app/README.md +12 -0
  369. examples/task_apps/verilog/task_app/__init__.py +1 -0
  370. examples/task_apps/verilog/task_app/grpo_verilog.py +1166 -0
  371. examples/task_apps/verilog/task_app/grpo_verilog_task_app.py +145 -0
  372. examples/task_apps/verilog/tests/__init__.py +4 -0
  373. examples/task_apps/verilog/tests/conftest.py +115 -0
  374. examples/task_apps/verilog/tests/integration/__init__.py +4 -0
  375. examples/task_apps/verilog/tests/integration/test_verilog_eval.py +181 -0
  376. examples/task_apps/verilog/tests/integration/test_verilog_rollout.py +55 -0
  377. examples/task_apps/verilog/tests/unit/__init__.py +4 -0
  378. examples/task_apps/verilog/tests/unit/test_verilog_scoring.py +118 -0
  379. examples/tunnel_gepa_banking77/README.md +106 -0
  380. examples/tunnel_gepa_banking77/banking77_gepa_tunnel.toml +95 -0
  381. examples/tunnel_gepa_banking77/keep_tunnel_running.py +60 -0
  382. examples/tunnel_gepa_banking77/run_gepa_with_tunnel.sh +226 -0
  383. examples/vlm/PROPOSAL.md +53 -0
  384. examples/vlm/README.md +68 -0
  385. examples/vlm/configs/crafter_vlm_gpt4o.toml +49 -0
  386. examples/vlm/crafter_image_only_agent.py +207 -0
  387. examples/vlm/crafter_openai_vlm_agent.py +275 -0
  388. examples/vlm/filter_image_rows.py +63 -0
  389. examples/vlm/run_crafter_vlm_benchmark.py +316 -0
  390. examples/warming_up_to_rl/_utils.py +92 -0
  391. examples/warming_up_to_rl/analyze_trace_db.py +422 -0
  392. examples/warming_up_to_rl/configs/crafter_fft.toml +53 -0
  393. examples/warming_up_to_rl/configs/crafter_fft_4b.toml +54 -0
  394. examples/warming_up_to_rl/configs/eval_fft_qwen4b.toml +22 -0
  395. examples/warming_up_to_rl/configs/eval_groq_qwen32b.toml +15 -0
  396. examples/warming_up_to_rl/configs/eval_modal_qwen4b.toml +24 -0
  397. examples/warming_up_to_rl/configs/eval_stepwise_complex.toml +35 -0
  398. examples/warming_up_to_rl/configs/eval_stepwise_consistent.toml +26 -0
  399. examples/warming_up_to_rl/configs/eval_stepwise_per_achievement.toml +36 -0
  400. examples/warming_up_to_rl/configs/eval_stepwise_simple.toml +32 -0
  401. examples/warming_up_to_rl/configs/rl_from_base_qwen4b.toml +85 -0
  402. examples/warming_up_to_rl/configs/rl_from_ft.toml +58 -0
  403. examples/warming_up_to_rl/export_trace_sft.py +837 -0
  404. examples/warming_up_to_rl/groq_test.py +97 -0
  405. examples/warming_up_to_rl/manage_secrets.py +131 -0
  406. examples/warming_up_to_rl/old/event_rewards.md +234 -0
  407. examples/warming_up_to_rl/old/notes.md +73 -0
  408. examples/warming_up_to_rl/readme.md +110 -0
  409. examples/warming_up_to_rl/run_eval.py +736 -0
  410. examples/warming_up_to_rl/run_fft_and_save.py +380 -0
  411. examples/warming_up_to_rl/run_local_rollout.py +239 -0
  412. examples/warming_up_to_rl/run_local_rollout_modal.py +248 -0
  413. examples/warming_up_to_rl/run_local_rollout_parallel.py +405 -0
  414. examples/warming_up_to_rl/run_local_rollout_traced.py +477 -0
  415. examples/warming_up_to_rl/run_rl_and_save.py +124 -0
  416. examples/warming_up_to_rl/run_rollout_remote.py +156 -0
  417. examples/warming_up_to_rl/task_app/README.md +42 -0
  418. examples/warming_up_to_rl/task_app/grpo_crafter.py +876 -0
  419. examples/warming_up_to_rl/task_app/grpo_crafter_task_app.py +135 -0
  420. examples/warming_up_to_rl/task_app/synth_envs_hosted/README.md +173 -0
  421. examples/warming_up_to_rl/task_app/synth_envs_hosted/__init__.py +5 -0
  422. examples/warming_up_to_rl/task_app/synth_envs_hosted/branching.py +143 -0
  423. examples/warming_up_to_rl/task_app/synth_envs_hosted/environment_routes.py +1226 -0
  424. examples/warming_up_to_rl/task_app/synth_envs_hosted/envs/__init__.py +1 -0
  425. examples/warming_up_to_rl/task_app/synth_envs_hosted/envs/crafter/__init__.py +6 -0
  426. examples/warming_up_to_rl/task_app/synth_envs_hosted/envs/crafter/app.py +1 -0
  427. examples/warming_up_to_rl/task_app/synth_envs_hosted/envs/crafter/environment.py +522 -0
  428. examples/warming_up_to_rl/task_app/synth_envs_hosted/envs/crafter/policy.py +454 -0
  429. examples/warming_up_to_rl/task_app/synth_envs_hosted/envs/crafter/react_agent.py +108 -0
  430. examples/warming_up_to_rl/task_app/synth_envs_hosted/envs/crafter/shared.py +305 -0
  431. examples/warming_up_to_rl/task_app/synth_envs_hosted/envs/crafter/tools.py +47 -0
  432. examples/warming_up_to_rl/task_app/synth_envs_hosted/hosted_app.py +253 -0
  433. examples/warming_up_to_rl/task_app/synth_envs_hosted/inference/__init__.py +5 -0
  434. examples/warming_up_to_rl/task_app/synth_envs_hosted/inference/openai_client.py +729 -0
  435. examples/warming_up_to_rl/task_app/synth_envs_hosted/main.py +100 -0
  436. examples/warming_up_to_rl/task_app/synth_envs_hosted/policy_routes.py +1114 -0
  437. examples/warming_up_to_rl/task_app/synth_envs_hosted/registry.py +195 -0
  438. examples/warming_up_to_rl/task_app/synth_envs_hosted/rollout.py +1891 -0
  439. examples/warming_up_to_rl/task_app/synth_envs_hosted/storage/__init__.py +5 -0
  440. examples/warming_up_to_rl/task_app/synth_envs_hosted/storage/volume.py +211 -0
  441. examples/warming_up_to_rl/task_app/synth_envs_hosted/test_agents.py +161 -0
  442. examples/warming_up_to_rl/task_app/synth_envs_hosted/test_service.py +137 -0
  443. examples/warming_up_to_rl/task_app/synth_envs_hosted/utils.py +129 -0
  444. examples/workflows/math_rl/configs/eval_base_qwen.toml +15 -0
  445. examples/workflows/math_rl/configs/eval_rl_qwen.toml +11 -0
  446. examples/workflows/math_rl/configs/rl_from_base_qwen.toml +62 -0
  447. examples/workflows/math_rl/configs/rl_from_base_qwen17.toml +80 -0
  448. examples/workflows/math_rl/configs/rl_from_ft_qwen.toml +35 -0
  449. examples/workflows/math_rl/download_dataset.py +80 -0
  450. examples/workflows/math_rl/run_eval.py +436 -0
  451. examples/workflows/math_rl/run_rl_and_save.py +111 -0
  452. synth_ai/__init__.py +47 -23
  453. synth_ai/_utils/__init__.py +47 -0
  454. synth_ai/_utils/base_url.py +10 -0
  455. synth_ai/_utils/http.py +10 -0
  456. synth_ai/_utils/prompts.py +10 -0
  457. synth_ai/_utils/task_app_state.py +12 -0
  458. synth_ai/_utils/user_config.py +10 -0
  459. synth_ai/api/models/supported.py +514 -0
  460. synth_ai/api/train/__init__.py +60 -2
  461. synth_ai/api/train/builders.py +347 -39
  462. synth_ai/api/train/cli.py +895 -160
  463. synth_ai/api/train/config_finder.py +103 -25
  464. synth_ai/api/train/configs/__init__.py +65 -0
  465. synth_ai/api/train/configs/prompt_learning.py +496 -0
  466. synth_ai/api/train/configs/rl.py +188 -0
  467. synth_ai/api/train/configs/sft.py +99 -0
  468. synth_ai/api/train/configs/shared.py +81 -0
  469. synth_ai/api/train/env_resolver.py +70 -20
  470. synth_ai/api/train/pollers.py +29 -4
  471. synth_ai/api/train/prompt_learning.py +425 -0
  472. synth_ai/api/train/sft.py +390 -0
  473. synth_ai/api/train/supported_algos.py +147 -0
  474. synth_ai/api/train/task_app.py +6 -4
  475. synth_ai/api/train/utils.py +64 -52
  476. synth_ai/api/train/validators.py +1117 -0
  477. synth_ai/api/tunnel.py +49 -0
  478. synth_ai/auth/credentials.py +94 -0
  479. synth_ai/baseline/__init__.py +25 -0
  480. synth_ai/baseline/config.py +209 -0
  481. synth_ai/baseline/discovery.py +214 -0
  482. synth_ai/baseline/execution.py +146 -0
  483. synth_ai/cfgs.py +227 -0
  484. synth_ai/cli/__init__.py +85 -63
  485. synth_ai/cli/_modal_wrapper.py +31 -0
  486. synth_ai/cli/_storage.py +20 -0
  487. synth_ai/cli/_typer_patch.py +47 -0
  488. synth_ai/cli/_validate_task_app.py +29 -0
  489. synth_ai/cli/balance.py +16 -4
  490. synth_ai/cli/calc.py +36 -21
  491. synth_ai/cli/claude.py +70 -0
  492. synth_ai/cli/codex.py +267 -0
  493. synth_ai/cli/commands/__init__.py +18 -0
  494. synth_ai/cli/commands/baseline/__init__.py +12 -0
  495. synth_ai/cli/commands/baseline/core.py +637 -0
  496. synth_ai/cli/commands/baseline/list.py +93 -0
  497. synth_ai/cli/commands/demo/__init__.py +6 -0
  498. synth_ai/cli/commands/demo/core.py +163 -0
  499. synth_ai/cli/commands/eval/__init__.py +19 -0
  500. synth_ai/cli/commands/eval/core.py +1112 -0
  501. synth_ai/cli/commands/eval/errors.py +81 -0
  502. synth_ai/cli/commands/eval/validation.py +133 -0
  503. synth_ai/cli/commands/filter/__init__.py +12 -0
  504. synth_ai/cli/commands/filter/core.py +424 -0
  505. synth_ai/cli/commands/filter/errors.py +55 -0
  506. synth_ai/cli/commands/filter/validation.py +77 -0
  507. synth_ai/cli/commands/help/__init__.py +185 -0
  508. synth_ai/cli/commands/help/core.py +72 -0
  509. synth_ai/cli/commands/smoke/__init__.py +7 -0
  510. synth_ai/cli/commands/smoke/core.py +1437 -0
  511. synth_ai/cli/commands/status/__init__.py +66 -0
  512. synth_ai/cli/commands/status/client.py +192 -0
  513. synth_ai/cli/commands/status/config.py +92 -0
  514. synth_ai/cli/commands/status/errors.py +20 -0
  515. synth_ai/cli/commands/status/formatters.py +164 -0
  516. synth_ai/cli/commands/status/subcommands/__init__.py +9 -0
  517. synth_ai/cli/commands/status/subcommands/files.py +79 -0
  518. synth_ai/cli/commands/status/subcommands/jobs.py +334 -0
  519. synth_ai/cli/commands/status/subcommands/models.py +79 -0
  520. synth_ai/cli/commands/status/subcommands/pricing.py +22 -0
  521. synth_ai/cli/commands/status/subcommands/runs.py +81 -0
  522. synth_ai/cli/commands/status/subcommands/session.py +183 -0
  523. synth_ai/cli/commands/status/subcommands/summary.py +47 -0
  524. synth_ai/cli/commands/status/subcommands/usage.py +203 -0
  525. synth_ai/cli/commands/status/utils.py +114 -0
  526. synth_ai/cli/commands/train/__init__.py +53 -0
  527. synth_ai/cli/commands/train/core.py +21 -0
  528. synth_ai/cli/commands/train/errors.py +117 -0
  529. synth_ai/cli/commands/train/judge_schemas.py +200 -0
  530. synth_ai/cli/commands/train/judge_validation.py +305 -0
  531. synth_ai/cli/commands/train/validation.py +386 -0
  532. synth_ai/cli/demo.py +32 -140
  533. synth_ai/cli/deploy.py +233 -0
  534. synth_ai/cli/eval/__init__.py +36 -0
  535. synth_ai/cli/eval/core.py +5 -0
  536. synth_ai/cli/eval/errors.py +31 -0
  537. synth_ai/cli/eval/validation.py +5 -0
  538. synth_ai/cli/filter/__init__.py +28 -0
  539. synth_ai/cli/filter/core.py +5 -0
  540. synth_ai/cli/filter/errors.py +23 -0
  541. synth_ai/cli/filter/validation.py +5 -0
  542. synth_ai/cli/legacy_root_backup.py +28 -22
  543. synth_ai/cli/lib/__init__.py +10 -0
  544. synth_ai/cli/lib/task_app_discovery.py +7 -0
  545. synth_ai/cli/lib/task_app_env.py +518 -0
  546. synth_ai/cli/mcp.py +34 -0
  547. synth_ai/cli/modal_serve/__init__.py +12 -0
  548. synth_ai/cli/modal_serve/core.py +14 -0
  549. synth_ai/cli/modal_serve/errors.py +8 -0
  550. synth_ai/cli/modal_serve/validation.py +11 -0
  551. synth_ai/cli/opencode.py +256 -0
  552. synth_ai/cli/recent.py +13 -7
  553. synth_ai/cli/rl_demo.py +156 -116
  554. synth_ai/cli/root.py +131 -132
  555. synth_ai/cli/serve/__init__.py +12 -0
  556. synth_ai/cli/serve/core.py +14 -0
  557. synth_ai/cli/serve/errors.py +8 -0
  558. synth_ai/cli/serve/validation.py +11 -0
  559. synth_ai/cli/setup.py +49 -0
  560. synth_ai/cli/status.py +7 -125
  561. synth_ai/cli/task_app_deploy.py +7 -0
  562. synth_ai/cli/task_app_list.py +25 -0
  563. synth_ai/cli/task_app_modal_serve.py +11 -0
  564. synth_ai/cli/task_app_serve.py +11 -0
  565. synth_ai/cli/task_apps.py +2284 -257
  566. synth_ai/cli/traces.py +9 -5
  567. synth_ai/cli/train/__init__.py +12 -0
  568. synth_ai/cli/train/core.py +21 -0
  569. synth_ai/cli/train/errors.py +8 -0
  570. synth_ai/cli/train/validation.py +24 -0
  571. synth_ai/cli/train.py +5 -0
  572. synth_ai/cli/turso.py +73 -0
  573. synth_ai/cli/watch.py +13 -18
  574. synth_ai/demos/__init__.py +10 -0
  575. synth_ai/demos/core/__init__.py +28 -1
  576. synth_ai/demos/core/cli.py +579 -291
  577. synth_ai/demos/crafter/__init__.py +1 -0
  578. synth_ai/demos/crafter/crafter_fft_4b.toml +55 -0
  579. synth_ai/demos/crafter/grpo_crafter_task_app.py +185 -0
  580. synth_ai/demos/crafter/rl_from_base_qwen4b.toml +74 -0
  581. synth_ai/demos/demo_registry.py +176 -0
  582. synth_ai/demos/demo_task_apps/__init__.py +3 -3
  583. synth_ai/demos/demo_task_apps/core.py +64 -28
  584. synth_ai/demos/demo_task_apps/crafter/__init__.py +1 -0
  585. synth_ai/demos/demo_task_apps/crafter/configs/crafter_fft_4b.toml +53 -0
  586. synth_ai/demos/demo_task_apps/crafter/configs/rl_from_base_qwen4b.toml +73 -0
  587. synth_ai/demos/demo_task_apps/crafter/grpo_crafter_task_app.py +184 -0
  588. synth_ai/demos/demo_task_apps/math/_common.py +1 -2
  589. synth_ai/demos/demo_task_apps/math/app.py +2 -1
  590. synth_ai/demos/demo_task_apps/math/deploy_modal.py +3 -6
  591. synth_ai/demos/demo_task_apps/math/modal_task_app.py +185 -83
  592. synth_ai/demos/demo_task_apps/math/task_app_entry.py +0 -2
  593. synth_ai/demos/math/__init__.py +1 -0
  594. synth_ai/demos/math/_common.py +16 -0
  595. synth_ai/demos/math/app.py +38 -0
  596. synth_ai/demos/math/config.toml +76 -0
  597. synth_ai/demos/math/deploy_modal.py +54 -0
  598. synth_ai/demos/math/modal_task_app.py +703 -0
  599. synth_ai/demos/math/task_app_entry.py +51 -0
  600. synth_ai/environments/environment/core.py +7 -1
  601. synth_ai/environments/examples/bandit/engine.py +12 -5
  602. synth_ai/environments/examples/bandit/environment.py +0 -1
  603. synth_ai/environments/examples/bandit/taskset.py +4 -4
  604. synth_ai/environments/examples/crafter_classic/engine_deterministic_patch.py +7 -4
  605. synth_ai/environments/examples/crafter_classic/engine_serialization_patch_v3.py +9 -5
  606. synth_ai/environments/examples/crafter_classic/environment.py +93 -2
  607. synth_ai/environments/examples/crafter_classic/world_config_patch_simple.py +4 -3
  608. synth_ai/environments/examples/enron/engine.py +7 -2
  609. synth_ai/environments/examples/enron/environment.py +68 -0
  610. synth_ai/environments/examples/red/engine.py +60 -12
  611. synth_ai/environments/examples/red/engine_helpers/memory_map.py +7 -0
  612. synth_ai/environments/examples/red/engine_helpers/reward_components.py +151 -179
  613. synth_ai/environments/examples/red/engine_helpers/reward_library/pallet_town_progression.py +477 -0
  614. synth_ai/environments/examples/red/engine_helpers/state_extraction.py +32 -0
  615. synth_ai/environments/examples/red/environment.py +86 -0
  616. synth_ai/environments/examples/red/trace_hooks_v3.py +168 -0
  617. synth_ai/environments/examples/sokoban/taskset.py +116 -0
  618. synth_ai/environments/examples/verilog/engine.py +104 -12
  619. synth_ai/environments/examples/wordle/environment.py +0 -1
  620. synth_ai/environments/reproducibility/tree.py +5 -6
  621. synth_ai/environments/service/app.py +11 -12
  622. synth_ai/environments/service/core_routes.py +10 -9
  623. synth_ai/environments/stateful/engine.py +1 -1
  624. synth_ai/environments/tasks/core.py +1 -0
  625. synth_ai/environments/tasks/filters.py +5 -6
  626. synth_ai/environments/tasks/utils.py +4 -5
  627. synth_ai/evals/__init__.py +15 -0
  628. synth_ai/evals/base.py +14 -5
  629. synth_ai/evals/client.py +82 -0
  630. synth_ai/evals/types.py +42 -0
  631. synth_ai/http.py +8 -22
  632. synth_ai/http_client.py +45 -12
  633. synth_ai/inference/__init__.py +0 -2
  634. synth_ai/inference/client.py +21 -7
  635. synth_ai/jobs/client.py +129 -80
  636. synth_ai/judge_schemas.py +127 -0
  637. synth_ai/learning/__init__.py +51 -6
  638. synth_ai/learning/algorithms.py +14 -0
  639. synth_ai/learning/client.py +122 -30
  640. synth_ai/learning/config.py +2 -40
  641. synth_ai/learning/constants.py +0 -2
  642. synth_ai/learning/ft_client.py +4 -56
  643. synth_ai/learning/health.py +14 -8
  644. synth_ai/learning/jobs.py +43 -47
  645. synth_ai/learning/prompt_learning_client.py +276 -0
  646. synth_ai/learning/prompt_learning_types.py +185 -0
  647. synth_ai/{rl → learning/rl}/__init__.py +14 -5
  648. synth_ai/learning/rl/client.py +269 -0
  649. synth_ai/learning/rl/config.py +31 -0
  650. synth_ai/{rl → learning/rl}/contracts.py +5 -10
  651. synth_ai/{rl → learning/rl}/env_keys.py +45 -16
  652. synth_ai/learning/rl/secrets.py +13 -0
  653. synth_ai/learning/rl_client.py +2 -253
  654. synth_ai/learning/sft/__init__.py +29 -0
  655. synth_ai/learning/sft/client.py +68 -0
  656. synth_ai/learning/sft/config.py +270 -0
  657. synth_ai/learning/sft/data.py +698 -0
  658. synth_ai/learning/sse.py +25 -26
  659. synth_ai/learning/validators.py +29 -25
  660. synth_ai/mcp/__init__.py +5 -0
  661. synth_ai/mcp/__main__.py +8 -0
  662. synth_ai/mcp/main.py +254 -0
  663. synth_ai/mcp/setup.py +100 -0
  664. synth_ai/modal.py +257 -0
  665. synth_ai/pricing/__init__.py +3 -0
  666. synth_ai/pricing/model_pricing.py +64 -0
  667. synth_ai/session/__init__.py +75 -0
  668. synth_ai/session/client.py +383 -0
  669. synth_ai/session/constants.py +63 -0
  670. synth_ai/session/exceptions.py +105 -0
  671. synth_ai/session/manager.py +139 -0
  672. synth_ai/session/models.py +89 -0
  673. synth_ai/session/query.py +110 -0
  674. synth_ai/spec/__init__.py +46 -0
  675. synth_ai/spec/dataclasses.py +149 -0
  676. synth_ai/spec/loader.py +144 -0
  677. synth_ai/spec/serializer.py +199 -0
  678. synth_ai/spec/validation.py +250 -0
  679. synth_ai/streaming/__init__.py +29 -0
  680. synth_ai/streaming/config.py +94 -0
  681. synth_ai/streaming/handlers.py +589 -0
  682. synth_ai/streaming/streamer.py +320 -0
  683. synth_ai/streaming/types.py +95 -0
  684. synth_ai/task/__init__.py +50 -30
  685. synth_ai/task/apps/__init__.py +63 -19
  686. synth_ai/task/auth.py +35 -23
  687. synth_ai/task/client.py +15 -13
  688. synth_ai/task/config.py +261 -0
  689. synth_ai/task/contracts.py +165 -64
  690. synth_ai/task/datasets.py +9 -6
  691. synth_ai/task/errors.py +11 -10
  692. synth_ai/task/health.py +17 -11
  693. synth_ai/task/inference_api.py +101 -0
  694. synth_ai/task/json.py +58 -24
  695. synth_ai/task/proxy.py +59 -66
  696. synth_ai/task/rubrics/__init__.py +55 -0
  697. synth_ai/task/rubrics/loaders.py +156 -0
  698. synth_ai/task/rubrics/models.py +57 -0
  699. synth_ai/task/rubrics/scoring.py +116 -0
  700. synth_ai/task/rubrics/strict.py +149 -0
  701. synth_ai/task/rubrics.py +22 -15
  702. synth_ai/task/server.py +65 -31
  703. synth_ai/task/trace_correlation_helpers.py +328 -0
  704. synth_ai/task/tracing_utils.py +44 -28
  705. synth_ai/task/validators.py +449 -6
  706. synth_ai/task/vendors.py +5 -7
  707. synth_ai/tracing_v3/__init__.py +4 -0
  708. synth_ai/tracing_v3/abstractions.py +21 -4
  709. synth_ai/tracing_v3/config.py +167 -22
  710. synth_ai/tracing_v3/constants.py +21 -0
  711. synth_ai/tracing_v3/db_config.py +42 -29
  712. synth_ai/tracing_v3/decorators.py +80 -45
  713. synth_ai/tracing_v3/examples/basic_usage.py +15 -9
  714. synth_ai/tracing_v3/hooks.py +6 -4
  715. synth_ai/tracing_v3/llm_call_record_helpers.py +161 -61
  716. synth_ai/tracing_v3/migration_helper.py +1 -2
  717. synth_ai/tracing_v3/replica_sync.py +12 -7
  718. synth_ai/tracing_v3/serialization.py +130 -0
  719. synth_ai/tracing_v3/session_tracer.py +73 -16
  720. synth_ai/tracing_v3/storage/base.py +89 -1
  721. synth_ai/tracing_v3/storage/config.py +63 -16
  722. synth_ai/tracing_v3/storage/factory.py +11 -9
  723. synth_ai/tracing_v3/storage/utils.py +15 -11
  724. synth_ai/tracing_v3/trace_utils.py +317 -0
  725. synth_ai/tracing_v3/turso/__init__.py +8 -21
  726. synth_ai/tracing_v3/turso/daemon.py +123 -15
  727. synth_ai/tracing_v3/turso/models.py +5 -2
  728. synth_ai/tracing_v3/turso/native_manager.py +1293 -0
  729. synth_ai/tracing_v3/utils.py +5 -4
  730. synth_ai/tunnel.py +143 -0
  731. synth_ai/tunnel_deploy.py +278 -0
  732. synth_ai/types.py +8 -0
  733. synth_ai/urls.py +11 -0
  734. synth_ai/utils/__init__.py +166 -0
  735. synth_ai/utils/agents.py +74 -0
  736. synth_ai/utils/apps.py +152 -0
  737. synth_ai/utils/base_url.py +94 -0
  738. synth_ai/utils/bin.py +39 -0
  739. synth_ai/utils/claude.py +36 -0
  740. synth_ai/utils/cli.py +284 -0
  741. synth_ai/utils/config.py +81 -0
  742. synth_ai/utils/env.py +346 -0
  743. synth_ai/utils/errors.py +85 -0
  744. synth_ai/utils/http.py +172 -0
  745. synth_ai/utils/json.py +72 -0
  746. synth_ai/utils/log_filter.py +99 -0
  747. synth_ai/utils/logging.py +198 -0
  748. synth_ai/utils/modal.py +299 -0
  749. synth_ai/utils/paths.py +95 -0
  750. synth_ai/utils/process.py +233 -0
  751. synth_ai/utils/prompts.py +39 -0
  752. synth_ai/utils/sqld.py +122 -0
  753. synth_ai/utils/ssl.py +25 -0
  754. synth_ai/utils/task_app_discovery.py +882 -0
  755. synth_ai/utils/task_app_env.py +186 -0
  756. synth_ai/utils/task_app_state.py +318 -0
  757. synth_ai/utils/tunnel/__init__.py +12 -0
  758. synth_ai/utils/tunnel/config.py +55 -0
  759. synth_ai/utils/user_config.py +137 -0
  760. synth_ai/uvicorn.py +77 -0
  761. synth_ai-0.2.23.dev3.dist-info/METADATA +357 -0
  762. synth_ai-0.2.23.dev3.dist-info/RECORD +983 -0
  763. {synth_ai-0.2.9.dev0.dist-info → synth_ai-0.2.23.dev3.dist-info}/entry_points.txt +0 -1
  764. {synth_ai-0.2.9.dev0.dist-info → synth_ai-0.2.23.dev3.dist-info}/top_level.txt +1 -0
  765. synth_ai/cli/man.py +0 -106
  766. synth_ai/core/experiment.py +0 -15
  767. synth_ai/core/system.py +0 -15
  768. synth_ai/demo_registry.py +0 -258
  769. synth_ai/environments/examples/sokoban/units/astar_common.py +0 -95
  770. synth_ai/experimental/synth_oss.py +0 -446
  771. synth_ai/handshake.py +0 -107
  772. synth_ai/install_sqld.sh +0 -40
  773. synth_ai/learning/offline/dpo.py +0 -0
  774. synth_ai/learning/offline/providers.py +0 -7
  775. synth_ai/learning/offline/sft.py +0 -0
  776. synth_ai/learning/offline/shared.py +0 -0
  777. synth_ai/learning/online/grpo.py +0 -0
  778. synth_ai/learning/online/irft.py +0 -0
  779. synth_ai/learning/prompts/banking77_injection_eval.py +0 -168
  780. synth_ai/learning/prompts/gepa.py +0 -0
  781. synth_ai/learning/prompts/hello_world_in_context_injection_ex.py +0 -213
  782. synth_ai/learning/prompts/mipro.py +0 -289
  783. synth_ai/learning/prompts/random_search.py +0 -246
  784. synth_ai/learning/prompts/run_mipro_banking77.py +0 -172
  785. synth_ai/learning/prompts/run_random_search_banking77.py +0 -324
  786. synth_ai/lm/__init__.py +0 -51
  787. synth_ai/lm/caching/constants.py +0 -6
  788. synth_ai/lm/caching/dbs.py +0 -0
  789. synth_ai/lm/caching/ephemeral.py +0 -102
  790. synth_ai/lm/caching/handler.py +0 -137
  791. synth_ai/lm/caching/initialize.py +0 -11
  792. synth_ai/lm/caching/persistent.py +0 -114
  793. synth_ai/lm/config.py +0 -110
  794. synth_ai/lm/constants.py +0 -32
  795. synth_ai/lm/core/__init__.py +0 -8
  796. synth_ai/lm/core/all.py +0 -73
  797. synth_ai/lm/core/exceptions.py +0 -7
  798. synth_ai/lm/core/main.py +0 -319
  799. synth_ai/lm/core/main_v3.py +0 -594
  800. synth_ai/lm/core/synth_models.py +0 -48
  801. synth_ai/lm/core/vendor_clients.py +0 -188
  802. synth_ai/lm/cost/monitor.py +0 -1
  803. synth_ai/lm/cost/statefulness.py +0 -1
  804. synth_ai/lm/injection.py +0 -80
  805. synth_ai/lm/overrides.py +0 -206
  806. synth_ai/lm/provider_support/__init__.py +0 -8
  807. synth_ai/lm/provider_support/anthropic.py +0 -972
  808. synth_ai/lm/provider_support/openai.py +0 -1139
  809. synth_ai/lm/provider_support/suppress_logging.py +0 -31
  810. synth_ai/lm/structured_outputs/handler.py +0 -440
  811. synth_ai/lm/structured_outputs/inject.py +0 -297
  812. synth_ai/lm/structured_outputs/rehabilitate.py +0 -185
  813. synth_ai/lm/tools/__init__.py +0 -3
  814. synth_ai/lm/tools/base.py +0 -172
  815. synth_ai/lm/unified_interface.py +0 -202
  816. synth_ai/lm/vendors/base.py +0 -81
  817. synth_ai/lm/vendors/core/anthropic_api.py +0 -387
  818. synth_ai/lm/vendors/core/gemini_api.py +0 -292
  819. synth_ai/lm/vendors/core/mistral_api.py +0 -322
  820. synth_ai/lm/vendors/core/openai_api.py +0 -225
  821. synth_ai/lm/vendors/core/synth_dev_api.py +0 -0
  822. synth_ai/lm/vendors/local/ollama.py +0 -0
  823. synth_ai/lm/vendors/openai_standard.py +0 -780
  824. synth_ai/lm/vendors/openai_standard_responses.py +0 -256
  825. synth_ai/lm/vendors/retries.py +0 -22
  826. synth_ai/lm/vendors/supported/custom_endpoint.py +0 -417
  827. synth_ai/lm/vendors/supported/deepseek.py +0 -69
  828. synth_ai/lm/vendors/supported/grok.py +0 -75
  829. synth_ai/lm/vendors/supported/groq.py +0 -16
  830. synth_ai/lm/vendors/supported/ollama.py +0 -15
  831. synth_ai/lm/vendors/supported/openrouter.py +0 -74
  832. synth_ai/lm/vendors/supported/together.py +0 -11
  833. synth_ai/lm/vendors/synth_client.py +0 -808
  834. synth_ai/lm/warmup.py +0 -186
  835. synth_ai/rl/secrets.py +0 -19
  836. synth_ai/scripts/verify_rewards.py +0 -100
  837. synth_ai/task/apps/grpo_crafter.py +0 -438
  838. synth_ai/tracing/__init__.py +0 -30
  839. synth_ai/tracing_v1/__init__.py +0 -33
  840. synth_ai/tracing_v3/turso/manager.py +0 -774
  841. synth_ai/v0/tracing/abstractions.py +0 -224
  842. synth_ai/v0/tracing/base_client.py +0 -91
  843. synth_ai/v0/tracing/client_manager.py +0 -131
  844. synth_ai/v0/tracing/config.py +0 -142
  845. synth_ai/v0/tracing/context.py +0 -146
  846. synth_ai/v0/tracing/decorators.py +0 -682
  847. synth_ai/v0/tracing/events/__init__.py +0 -0
  848. synth_ai/v0/tracing/events/manage.py +0 -147
  849. synth_ai/v0/tracing/events/scope.py +0 -86
  850. synth_ai/v0/tracing/events/store.py +0 -228
  851. synth_ai/v0/tracing/immediate_client.py +0 -151
  852. synth_ai/v0/tracing/local.py +0 -18
  853. synth_ai/v0/tracing/log_client_base.py +0 -73
  854. synth_ai/v0/tracing/retry_queue.py +0 -186
  855. synth_ai/v0/tracing/trackers.py +0 -515
  856. synth_ai/v0/tracing/upload.py +0 -512
  857. synth_ai/v0/tracing/utils.py +0 -9
  858. synth_ai/v0/tracing_v1/__init__.py +0 -16
  859. synth_ai/v0/tracing_v1/abstractions.py +0 -224
  860. synth_ai/v0/tracing_v1/base_client.py +0 -91
  861. synth_ai/v0/tracing_v1/client_manager.py +0 -131
  862. synth_ai/v0/tracing_v1/config.py +0 -142
  863. synth_ai/v0/tracing_v1/context.py +0 -146
  864. synth_ai/v0/tracing_v1/decorators.py +0 -703
  865. synth_ai/v0/tracing_v1/events/__init__.py +0 -0
  866. synth_ai/v0/tracing_v1/events/manage.py +0 -147
  867. synth_ai/v0/tracing_v1/events/scope.py +0 -86
  868. synth_ai/v0/tracing_v1/events/store.py +0 -228
  869. synth_ai/v0/tracing_v1/immediate_client.py +0 -151
  870. synth_ai/v0/tracing_v1/local.py +0 -18
  871. synth_ai/v0/tracing_v1/log_client_base.py +0 -73
  872. synth_ai/v0/tracing_v1/retry_queue.py +0 -186
  873. synth_ai/v0/tracing_v1/trackers.py +0 -515
  874. synth_ai/v0/tracing_v1/upload.py +0 -527
  875. synth_ai/v0/tracing_v1/utils.py +0 -9
  876. synth_ai/zyk/__init__.py +0 -30
  877. synth_ai-0.2.9.dev0.dist-info/METADATA +0 -131
  878. synth_ai-0.2.9.dev0.dist-info/RECORD +0 -444
  879. {synth_ai/lm/caching → examples/task_apps}/__init__.py +0 -0
  880. {synth_ai/lm/cost → examples/task_apps/crafter}/__init__.py +0 -0
  881. {synth_ai/lm/structured_outputs → examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/server}/__init__.py +0 -0
  882. {synth_ai/lm/vendors → examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/tests}/__init__.py +0 -0
  883. {synth_ai/lm/vendors/core → examples/task_apps/dev/pokemon_emerald/external/pokeagent-speedrun/utils}/__init__.py +0 -0
  884. {synth_ai/lm/vendors/local → examples/task_apps/math}/__init__.py +0 -0
  885. {synth_ai/lm/vendors/supported → examples/workflows}/__init__.py +0 -0
  886. {synth_ai/v0/tracing → examples/workflows/math_rl}/__init__.py +0 -0
  887. /synth_ai/{compound/cais.py → cli/__main__.py} +0 -0
  888. /synth_ai/{learning/filtering.py → py.typed} +0 -0
  889. {synth_ai-0.2.9.dev0.dist-info → synth_ai-0.2.23.dev3.dist-info}/WHEEL +0 -0
  890. {synth_ai-0.2.9.dev0.dist-info → synth_ai-0.2.23.dev3.dist-info}/licenses/LICENSE +0 -0
@@ -1,168 +0,0 @@
1
- """
2
- Banking77 in-context injection evals (async, not tests)
3
-
4
- Samples a handful of Banking77 prompts and evaluates multiple override
5
- contexts in parallel, printing simple accuracy for each.
6
-
7
- Usage
8
- - Keys in .env (GROQ_API_KEY, etc.)
9
- - Run: uv run -q python -m synth_ai.learning.prompts.banking77_injection_eval
10
- Optional env:
11
- - N_SAMPLES=20 (default)
12
- - MODEL=openai/gpt-oss-20b (default)
13
- - VENDOR=groq (default)
14
- """
15
-
16
- from __future__ import annotations
17
-
18
- import asyncio
19
- import os
20
- import random
21
- from typing import Any
22
-
23
- from datasets import load_dataset
24
- from dotenv import load_dotenv
25
- from synth_ai.lm.core.main_v3 import LM, build_messages
26
- from synth_ai.lm.overrides import LMOverridesContext
27
-
28
-
29
- async def classify_one(lm: LM, text: str, label_names: list[str]) -> str:
30
- labels_joined = ", ".join(label_names)
31
- system_message = (
32
- "You are an intent classifier for the Banking77 dataset. "
33
- "Given a customer message, respond with exactly one label from the list. "
34
- "Return only the label text with no extra words.\n\n"
35
- f"Valid labels: {labels_joined}"
36
- )
37
- user_message = f"Message: {text}\nLabel:"
38
- messages = build_messages(system_message, user_message, images_bytes=None, model_name=lm.model)
39
- resp = await lm.respond_async(messages=messages)
40
- return (resp.raw_response or "").strip()
41
-
42
-
43
- def choose_label(pred: str, label_names: list[str]) -> str:
44
- norm_pred = pred.strip().lower()
45
- label_lookup = {ln.lower(): ln for ln in label_names}
46
- mapped = label_lookup.get(norm_pred)
47
- if mapped is not None:
48
- return mapped
49
-
50
- # Fallback: choose the label with the highest naive token overlap
51
- def score(cand: str) -> int:
52
- c = cand.lower()
53
- return sum(1 for w in c.split() if w in norm_pred)
54
-
55
- return max(label_names, key=score)
56
-
57
-
58
- async def eval_context(
59
- lm: LM,
60
- items: list[tuple[str, str]],
61
- label_names: list[str],
62
- ctx_name: str,
63
- specs: list[dict[str, Any]],
64
- ) -> tuple[str, int, int]:
65
- correct = 0
66
- with LMOverridesContext(specs):
67
- tasks = [classify_one(lm, text, label_names) for text, _ in items]
68
- results = await asyncio.gather(*tasks, return_exceptions=True)
69
- for (text, gold), pred in zip(items, results, strict=False):
70
- if isinstance(pred, Exception):
71
- # Treat exceptions as incorrect
72
- continue
73
- mapped = choose_label(pred, label_names)
74
- correct += int(mapped == gold)
75
- return (ctx_name, correct, len(items))
76
-
77
-
78
- async def main() -> None:
79
- load_dotenv()
80
-
81
- n = int(os.getenv("N_SAMPLES", "20"))
82
- model = os.getenv("MODEL", "openai/gpt-oss-20b")
83
- vendor = os.getenv("VENDOR", "groq")
84
-
85
- lm = LM(model=model, vendor=vendor, temperature=0.0)
86
-
87
- print("Loading Banking77 dataset (split='test')...")
88
- ds = load_dataset("banking77", split="test")
89
- label_names: list[str] = ds.features["label"].names # type: ignore
90
-
91
- idxs = random.sample(range(len(ds)), k=min(n, len(ds)))
92
- items = [
93
- (ds[i]["text"], label_names[int(ds[i]["label"])]) # (text, gold_label)
94
- for i in idxs
95
- ]
96
-
97
- # Define a few override contexts to compare
98
- contexts: list[dict[str, Any]] = [
99
- {
100
- "name": "baseline (no overrides)",
101
- "overrides": [],
102
- },
103
- {
104
- "name": "nonsense prompt injection (expected worse)",
105
- "overrides": [
106
- {
107
- "match": {"contains": "", "role": "user"},
108
- "injection_rules": [
109
- # Heavily corrupt user text by replacing vowels
110
- {"find": "a", "replace": "x"},
111
- {"find": "e", "replace": "x"},
112
- {"find": "i", "replace": "x"},
113
- {"find": "o", "replace": "x"},
114
- {"find": "u", "replace": "x"},
115
- {"find": "A", "replace": "X"},
116
- {"find": "E", "replace": "X"},
117
- {"find": "I", "replace": "X"},
118
- {"find": "O", "replace": "X"},
119
- {"find": "U", "replace": "X"},
120
- ],
121
- }
122
- ],
123
- },
124
- {
125
- "name": "injection: atm->ATM, txn->transaction",
126
- "overrides": [
127
- {
128
- "match": {"contains": "atm", "role": "user"},
129
- "injection_rules": [
130
- {"find": "atm", "replace": "ATM"},
131
- {"find": "txn", "replace": "transaction"},
132
- ],
133
- }
134
- ],
135
- },
136
- {
137
- "name": "params: temperature=0.0",
138
- "overrides": [
139
- {"match": {"contains": ""}, "params": {"temperature": 0.0}},
140
- ],
141
- },
142
- {
143
- "name": "model override: 20b->120b",
144
- "overrides": [
145
- {"match": {"contains": ""}, "params": {"model": "openai/gpt-oss-120b"}},
146
- ],
147
- },
148
- ]
149
-
150
- print(f"\nEvaluating {len(contexts)} contexts on {len(items)} Banking77 samples (async)...")
151
-
152
- # Evaluate each context sequentially but batched (each context classifies in parallel)
153
- results: list[tuple[str, int, int]] = []
154
- for ctx in contexts:
155
- name = ctx["name"]
156
- specs = ctx["overrides"]
157
- print(f"Evaluating: {name} ...")
158
- res = await eval_context(lm, items, label_names, name, specs)
159
- results.append(res)
160
-
161
- print("\nResults:")
162
- for name, correct, total in results:
163
- acc = correct / total if total else 0.0
164
- print(f"- {name}: {correct}/{total} correct ({acc:.2%})")
165
-
166
-
167
- if __name__ == "__main__":
168
- asyncio.run(main())
File without changes
@@ -1,213 +0,0 @@
1
- """
2
- Hello World: Banking77 intent classification with in-context injection
3
-
4
- This script shows a minimal text-classification pipeline over the
5
- Hugging Face Banking77 dataset using the Synth LM interface. It also
6
- demonstrates a simple pre-send prompt-injection step as outlined in
7
- `synth_ai/learning/prompts/injection_plan.txt`.
8
-
9
- Notes
10
- - Network access is required to download the dataset and call the model.
11
- - Defaults to Groq with model `openai/gpt-oss-20b`.
12
- - Export your key: `export GROQ_API_KEY=...`
13
- - Override if needed: `export MODEL=openai/gpt-oss-20b VENDOR=groq`
14
-
15
- Run
16
- - `python -m synth_ai.learning.prompts.hello_world_in_context_injection_ex`
17
-
18
- What "in-context injection" means here
19
- - The script applies ordered substring replacements to the outgoing
20
- `messages` array before calling the model. This mirrors the algorithm
21
- described in `injection_plan.txt` without importing any non-existent
22
- helper yet. You can adapt `INJECTION_RULES` to your needs.
23
- """
24
-
25
- from __future__ import annotations
26
-
27
- import asyncio
28
- import os
29
- import random
30
-
31
- from datasets import load_dataset
32
-
33
- # Use the v3 LM class present in this repo
34
- from synth_ai.lm.core.main_v3 import LM, build_messages
35
-
36
- # Use Overrides context to demonstrate matching by content
37
- from synth_ai.lm.overrides import LMOverridesContext
38
- from synth_ai.tracing_v3.abstractions import LMCAISEvent
39
- from synth_ai.tracing_v3.session_tracer import SessionTracer
40
-
41
- INJECTION_RULES = [
42
- {"find": "accnt", "replace": "account"},
43
- {"find": "atm", "replace": "ATM"},
44
- {"find": "txn", "replace": "transaction"},
45
- ]
46
-
47
-
48
- async def classify_sample(lm: LM, text: str, label_names: list[str]) -> str:
49
- """Classify one Banking77 utterance and return the predicted label name."""
50
- labels_joined = ", ".join(label_names)
51
- system_message = (
52
- "You are an intent classifier for the Banking77 dataset. "
53
- "Given a customer message, respond with exactly one label from the list. "
54
- "Return only the label text with no extra words.\n\n"
55
- f"Valid labels: {labels_joined}"
56
- )
57
- user_message = f"Message: {text}\nLabel:"
58
-
59
- # Build canonical messages; injection will be applied inside the vendor via context
60
- messages = build_messages(system_message, user_message, images_bytes=None, model_name=lm.model)
61
- resp = await lm.respond_async(messages=messages)
62
- raw = (resp.raw_response or "").strip()
63
- return raw
64
-
65
-
66
- async def main() -> None:
67
- # Configurable model/provider via env, with sensible defaults
68
- # Default to Groq hosting `openai/gpt-oss-20b`
69
- model = os.getenv("MODEL", "openai/gpt-oss-20b")
70
- vendor = os.getenv("VENDOR", "groq")
71
-
72
- # Construct LM
73
- lm = LM(model=model, vendor=vendor, temperature=0.0)
74
-
75
- # Load Banking77 dataset
76
- # Columns: {"text": str, "label": int}; label names at ds.features["label"].names
77
- print("Loading Banking77 dataset (split='test')...")
78
- ds = load_dataset("banking77", split="test")
79
- label_names: list[str] = ds.features["label"].names # type: ignore
80
-
81
- # Sample a few items for a quick demo
82
- n = int(os.getenv("N_SAMPLES", "8"))
83
- idxs = random.sample(range(len(ds)), k=min(n, len(ds)))
84
-
85
- correct = 0
86
- # Apply overrides for all calls in this block (match by content)
87
- overrides = [
88
- {"match": {"contains": "atm", "role": "user"}, "injection_rules": INJECTION_RULES},
89
- {"match": {"contains": "refund"}, "params": {"temperature": 0.0}},
90
- ]
91
- with LMOverridesContext(overrides):
92
- for i, idx in enumerate(idxs, start=1):
93
- text: str = ds[idx]["text"] # type: ignore
94
- gold_label_idx: int = int(ds[idx]["label"]) # type: ignore
95
- gold_label = label_names[gold_label_idx]
96
-
97
- try:
98
- pred = await classify_sample(lm, text, label_names)
99
- except Exception as e:
100
- print(f"[{i}] Error calling model: {e}")
101
- break
102
-
103
- # Normalize and check exact match; if not exact, attempt a loose fallback
104
- norm_pred = pred.strip().lower()
105
- label_lookup = {ln.lower(): ln for ln in label_names}
106
- pred_label = label_lookup.get(norm_pred)
107
- if pred_label is None:
108
- # Fallback: pick the label with highest substring overlap (very naive)
109
- # This avoids extra deps; feel free to replace with a better matcher.
110
- def score(cand: str) -> int:
111
- c = cand.lower()
112
- return sum(1 for w in c.split() if w in norm_pred)
113
-
114
- pred_label = max(label_names, key=score)
115
-
116
- is_correct = pred_label == gold_label
117
- correct += int(is_correct)
118
- print(
119
- f"[{i}] text={text!r}\n gold={gold_label}\n pred={pred} -> mapped={pred_label} {'✅' if is_correct else '❌'}"
120
- )
121
-
122
- if idxs:
123
- acc = correct / len(idxs)
124
- print(f"\nSamples: {len(idxs)} | Correct: {correct} | Accuracy: {acc:.2%}")
125
-
126
- # ------------------------------
127
- # Integration tests (three paths)
128
- # ------------------------------
129
- print("\nRunning integration tests with in-context injection...")
130
- test_text = "I used the atm to withdraw cash."
131
-
132
- # 1) LM path with v3 tracing: verify substitution in traced messages
133
- tracer = SessionTracer()
134
- await tracer.start_session(metadata={"test": "lm_injection"})
135
- await tracer.start_timestep(step_id="lm_test")
136
- # Use a tracer-bound LM instance
137
- lm_traced = LM(model=model, vendor=vendor, temperature=0.0, session_tracer=tracer)
138
- with LMOverridesContext([{"match": {"contains": "atm"}, "injection_rules": INJECTION_RULES}]):
139
- _ = await classify_sample(lm_traced, test_text, label_names)
140
- # inspect trace
141
- events = [
142
- e
143
- for e in (tracer.current_session.event_history if tracer.current_session else [])
144
- if isinstance(e, LMCAISEvent)
145
- ]
146
- assert events, "No LMCAISEvent recorded by SessionTracer"
147
- cr = events[-1].call_records[0]
148
- traced_user = ""
149
- for m in cr.input_messages:
150
- if m.role == "user":
151
- for part in m.parts:
152
- if getattr(part, "type", None) == "text":
153
- traced_user += part.text or ""
154
- assert "ATM" in traced_user, f"Expected substitution in traced prompt; got: {traced_user!r}"
155
- print("LM path trace verified: substitution present in traced prompt.")
156
- await tracer.end_timestep()
157
- await tracer.end_session()
158
-
159
- # 2) OpenAI wrapper path (AsyncOpenAI to Groq): ensure apply_injection is active
160
- try:
161
- import synth_ai.lm.provider_support.openai as _synth_openai_patch # noqa: F401
162
- from openai import AsyncOpenAI
163
-
164
- base_url = os.getenv("OPENAI_BASE_URL", "https://api.groq.com/openai/v1")
165
- api_key = os.getenv("OPENAI_API_KEY") or os.getenv("GROQ_API_KEY") or ""
166
- client = AsyncOpenAI(base_url=base_url, api_key=api_key)
167
- messages = [
168
- {"role": "system", "content": "Echo user label."},
169
- {"role": "user", "content": f"Please classify: {test_text}"},
170
- ]
171
- with LMOverridesContext(
172
- [{"match": {"contains": "atm"}, "injection_rules": INJECTION_RULES}]
173
- ):
174
- _ = await client.chat.completions.create(
175
- model=model, messages=messages, temperature=0
176
- )
177
- # Not all models echo input; instead, verify that our injected expectation matches
178
- expected_user = messages[1]["content"].replace("atm", "ATM")
179
- if messages[1]["content"] == expected_user:
180
- print("OpenAI wrapper: input already normalized; skipping assertion.")
181
- else:
182
- print("OpenAI wrapper: sent message contains substitution expectation:", expected_user)
183
- except Exception as e:
184
- print("OpenAI wrapper test skipped due to error:", e)
185
-
186
- # 3) Anthropic wrapper path (AsyncClient): ensure apply_injection is active
187
- try:
188
- import anthropic
189
- import synth_ai.lm.provider_support.anthropic as _synth_anthropic_patch # noqa: F401
190
-
191
- a_model = os.getenv("ANTHROPIC_MODEL", "claude-3-5-haiku-20241022")
192
- a_key = os.getenv("ANTHROPIC_API_KEY")
193
- if a_key:
194
- a_client = anthropic.AsyncClient(api_key=a_key)
195
- with LMOverridesContext(
196
- [{"match": {"contains": "atm"}, "injection_rules": INJECTION_RULES}]
197
- ):
198
- _ = await a_client.messages.create(
199
- model=a_model,
200
- system="Echo user label.",
201
- max_tokens=64,
202
- temperature=0,
203
- messages=[{"role": "user", "content": [{"type": "text", "text": test_text}]}],
204
- )
205
- print("Anthropic wrapper call completed (cannot reliably assert echo).")
206
- else:
207
- print("Anthropic wrapper test skipped: ANTHROPIC_API_KEY not set.")
208
- except Exception as e:
209
- print("Anthropic wrapper test skipped due to error:", e)
210
-
211
-
212
- if __name__ == "__main__":
213
- asyncio.run(main())
@@ -1,289 +0,0 @@
1
- """
2
- MIPROv2-style prompt optimizer (modular, DSPy-inspired).
3
-
4
- This module provides a modular implementation of the MIPROv2 pseudocode from DSPy,
5
- adapted to a provider-agnostic "program" interface. The goal is to keep the
6
- bootstrapping and search process pluggable so it can be swapped for alternatives.
7
-
8
- Key ideas
9
- - Program adapter: unify how we set instructions/demos and run predictions.
10
- - Demo bootstrapping: gather high-confidence examples (by metric) as candidates.
11
- - Instruction proposals: generated by a prompt model from contextual summaries.
12
- - Search (placeholder): random/Bayesian-like search over (instructions × demos).
13
-
14
- Notes
15
- - The implementation is intentionally lightweight and dependency-free.
16
- - "BayesOpt" here is a placeholder randomized proposer that uses history; you
17
- can plug in a real optimizer later.
18
- """
19
-
20
- from __future__ import annotations
21
-
22
- import random
23
- from collections.abc import Callable, Sequence
24
- from dataclasses import dataclass, replace
25
- from typing import Any, Protocol
26
-
27
- # ---------------------------
28
- # Program adapter and protocols
29
- # ---------------------------
30
-
31
-
32
- class PredictProgram(Protocol):
33
- """Minimal protocol a program must satisfy for MIPRO.
34
-
35
- You can adapt your own pipeline to this by implementing these methods or
36
- by wrapping it with `ProgramAdapter` below.
37
- """
38
-
39
- def deepcopy(self) -> PredictProgram: ...
40
-
41
- def run(self, x: Any, *, model: Any | None = None) -> Any: ...
42
-
43
- def with_instructions(self, instructions: dict[str, str]) -> PredictProgram: ...
44
-
45
- def with_demos(self, demos: list[tuple[Any, Any]]) -> PredictProgram: ...
46
-
47
- @property
48
- def predictors(self) -> list[str]: ...
49
-
50
-
51
- @dataclass
52
- class ProgramAdapter:
53
- """Adapter that turns a set of callables/state into a `PredictProgram`.
54
-
55
- - run_fn: Callable[[x, model], y]
56
- - state: arbitrary dict; supports `instructions` and `demos` keys
57
- - predictors: list of predictor identifiers (e.g., names of prompt blocks)
58
- - set_instructions: Callable to update instructions (per predictor)
59
- - set_demos: Callable to update demos (global or per predictor)
60
- """
61
-
62
- run_fn: Callable[[Any, Any | None], Any]
63
- state: dict[str, Any]
64
- _predictors: list[str]
65
- set_instructions: Callable[[dict[str, str], dict[str, Any]], dict[str, Any]]
66
- set_demos: Callable[[list[tuple[Any, Any]], dict[str, Any]], dict[str, Any]]
67
-
68
- def deepcopy(self) -> ProgramAdapter:
69
- return replace(self, state={**self.state})
70
-
71
- def run(self, x: Any, *, model: Any | None = None) -> Any:
72
- return self.run_fn(x, model)
73
-
74
- def with_instructions(self, instructions: dict[str, str]) -> ProgramAdapter:
75
- new_state = self.set_instructions(instructions, {**self.state})
76
- return replace(self, state=new_state)
77
-
78
- def with_demos(self, demos: list[tuple[Any, Any]]) -> ProgramAdapter:
79
- new_state = self.set_demos(demos, {**self.state})
80
- return replace(self, state=new_state)
81
-
82
- @property
83
- def predictors(self) -> list[str]:
84
- return list(self._predictors)
85
-
86
-
87
- # ---------------------------
88
- # Utility helpers
89
- # ---------------------------
90
-
91
-
92
- def summarize_dataset(trainset: Sequence[tuple[Any, Any]], max_items: int = 50) -> str:
93
- n = len(trainset)
94
- ex = ", ".join(
95
- repr(trainset[i][0])[:40] for i in range(0, min(max_items, n), max(1, n // max_items or 1))
96
- )
97
- return f"Dataset size: {n}. Example inputs: {ex}"
98
-
99
-
100
- def summarize_program(prog: PredictProgram) -> str:
101
- return f"Program predictors: {prog.predictors}"
102
-
103
-
104
- def random_tip(rng: random.Random) -> str:
105
- tips = [
106
- "Be concise.",
107
- "Focus on the task definition.",
108
- "Use the provided examples as guidance.",
109
- "Avoid unnecessary verbosity.",
110
- ]
111
- return rng.choice(tips)
112
-
113
-
114
- def choose(items: Sequence[Any], rng: random.Random | None = None) -> Any:
115
- r = rng or random
116
- return r.choice(items)
117
-
118
-
119
- # ---------------------------
120
- # Evaluator
121
- # ---------------------------
122
-
123
-
124
- @dataclass
125
- class EvalResult:
126
- score: float
127
- subscores: list[float]
128
-
129
-
130
- def evaluate_program(
131
- program: PredictProgram, dataset: Sequence[tuple[Any, Any]], metric: Callable[[Any, Any], float]
132
- ) -> EvalResult:
133
- subs = []
134
- for x, y in dataset:
135
- yhat = program.run(x)
136
- subs.append(metric(yhat, y))
137
- return EvalResult(score=float(sum(subs)) / max(1, len(subs)), subscores=subs)
138
-
139
-
140
- # ---------------------------
141
- # MIPROv2 compile
142
- # ---------------------------
143
-
144
-
145
- def mipro_v2_compile(
146
- student: PredictProgram,
147
- trainset: Sequence[tuple[Any, Any]],
148
- valset: Sequence[tuple[Any, Any]],
149
- metric: Callable[[Any, Any], float],
150
- *,
151
- prompt_model: Any,
152
- task_model: Any,
153
- max_bootstrapped_demos: int = 8,
154
- max_labeled_demos: int = 4,
155
- num_candidates: int = 8,
156
- num_trials: int = 20,
157
- minibatch: bool = True,
158
- minibatch_size: int = 16,
159
- minibatch_full_eval_steps: int = 5,
160
- seed: int = 0,
161
- auto: str = "light",
162
- program_aware: bool = True,
163
- data_aware: bool = True,
164
- tip_aware: bool = True,
165
- fewshot_aware: bool = True,
166
- ) -> tuple[PredictProgram, list[dict[str, Any]]]:
167
- """MIPROv2-style optimizer.
168
-
169
- Arguments mirror the DSPy pseudocode but remain provider-agnostic. The
170
- `prompt_model` must expose `generate_instructions(ctx, k)`; the `student`
171
- program must implement the `PredictProgram` protocol.
172
- """
173
-
174
- rng = random.Random(seed)
175
- program = student.deepcopy()
176
-
177
- # Step 1: bootstrap few-shot example candidates
178
- demo_candidates: list[dict[str, Any]] = []
179
- for _ in range(num_candidates):
180
- boot: list[tuple[Any, Any]] = []
181
- # collect bootstrapped, self-consistent demos
182
- while len(boot) < max_bootstrapped_demos:
183
- x, y = rng.choice(trainset)
184
- yhat = program.run(x, model=task_model)
185
- if metric(yhat, y) == 1: # perfect match
186
- boot.append((x, y))
187
- labeled = rng.sample(list(trainset), k=min(max_labeled_demos, len(trainset)))
188
- demo_candidates.append({"boot": boot, "labeled": labeled})
189
-
190
- # Step 2: propose instruction candidates per predictor
191
- instr_candidates: dict[str, list[str]] = {}
192
- for pred in program.predictors or ["predictor"]:
193
- ctx: dict[str, Any] = {}
194
- if data_aware:
195
- ctx["dataset_summary"] = summarize_dataset(trainset)
196
- if program_aware:
197
- ctx["program_summary"] = summarize_program(program)
198
- if fewshot_aware and demo_candidates:
199
- ctx["examples"] = choose(demo_candidates, rng)
200
- if tip_aware:
201
- ctx["tip"] = random_tip(rng)
202
- cand = prompt_model.generate_instructions(ctx, k=num_candidates)
203
- instr_candidates[pred] = list(cand)
204
-
205
- # Step 3: Bayesian-optimization-like search (random proposer placeholder)
206
- history: list[tuple[dict[str, Any], float]] = []
207
- records: list[dict[str, Any]] = []
208
- best_score = -1.0
209
- best_cfg: dict[str, Any] | None = None
210
-
211
- def propose(history_: list[tuple[dict[str, Any], float]]) -> dict[str, Any]:
212
- # Placeholder: randomly sample from the cartesian product
213
- instructions = {pred: choose(instr_candidates[pred], rng) for pred in instr_candidates}
214
- demos = choose(demo_candidates, rng) if demo_candidates else None
215
- return {"instructions": instructions, "demo_set": demos}
216
-
217
- for t in range(1, num_trials + 1):
218
- theta = propose(history)
219
- program_t = program.with_instructions(theta["instructions"])
220
- if theta.get("demo_set") is not None:
221
- # Combine bootstrapped + labeled demos
222
- ds = theta["demo_set"]
223
- demo_set = list(ds.get("boot", [])) + list(ds.get("labeled", []))
224
- program_t = program_t.with_demos(demo_set)
225
-
226
- batch = (
227
- valset
228
- if not minibatch
229
- else random.sample(list(valset), k=min(minibatch_size, len(valset)))
230
- )
231
- batch_res = evaluate_program(program_t, batch, metric)
232
- s_t = batch_res.score
233
- history.append((theta, s_t))
234
- records.append(
235
- {
236
- "trial": t,
237
- "evaluation": "batch" if minibatch else "full",
238
- "score": s_t,
239
- "intervention": {
240
- "instructions": theta.get("instructions"),
241
- "demo_set": theta.get("demo_set"),
242
- },
243
- }
244
- )
245
-
246
- if (not minibatch) or (t % max(1, minibatch_full_eval_steps) == 0):
247
- full_res = evaluate_program(program_t, valset, metric)
248
- s_full = full_res.score
249
- if s_full > best_score:
250
- best_score = s_full
251
- best_cfg = theta
252
- records.append(
253
- {
254
- "trial": t,
255
- "evaluation": "full",
256
- "score": s_full,
257
- "intervention": {
258
- "instructions": theta.get("instructions"),
259
- "demo_set": theta.get("demo_set"),
260
- },
261
- }
262
- )
263
-
264
- if best_cfg is None:
265
- return program, records
266
-
267
- best_program = program.with_instructions(best_cfg["instructions"])
268
- if best_cfg.get("demo_set") is not None:
269
- ds = best_cfg["demo_set"]
270
- demo_set = list(ds.get("boot", [])) + list(ds.get("labeled", []))
271
- best_program = best_program.with_demos(demo_set)
272
- return best_program, records
273
-
274
-
275
- __all__ = [
276
- "PredictProgram",
277
- "ProgramAdapter",
278
- "evaluate_program",
279
- "mipro_v2_compile",
280
- ]
281
-
282
-
283
- class ExampleTwoStepDag:
284
- pass
285
-
286
-
287
- """
288
- A -> B
289
- """