decodingtrust-agent-sdk 0.1.0__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 (374) hide show
  1. agent/__init__.py +30 -0
  2. agent/claudesdk/__init__.py +8 -0
  3. agent/claudesdk/example.py +221 -0
  4. agent/claudesdk/src/__init__.py +8 -0
  5. agent/claudesdk/src/agent.py +400 -0
  6. agent/claudesdk/src/mcp_proxy.py +409 -0
  7. agent/claudesdk/src/utils.py +420 -0
  8. agent/googleadk/__init__.py +15 -0
  9. agent/googleadk/example.py +237 -0
  10. agent/googleadk/src/__init__.py +12 -0
  11. agent/googleadk/src/agent.py +401 -0
  12. agent/googleadk/src/mcp_wrapper.py +163 -0
  13. agent/googleadk/src/utils.py +602 -0
  14. agent/langchain/__init__.py +8 -0
  15. agent/langchain/example.py +213 -0
  16. agent/langchain/src/__init__.py +8 -0
  17. agent/langchain/src/agent.py +645 -0
  18. agent/langchain/src/utils.py +433 -0
  19. agent/openaisdk/__init__.py +17 -0
  20. agent/openaisdk/example.py +228 -0
  21. agent/openaisdk/src/__init__.py +12 -0
  22. agent/openaisdk/src/agent.py +491 -0
  23. agent/openaisdk/src/agent_wrapper.py +143 -0
  24. agent/openaisdk/src/mcp_wrapper.py +395 -0
  25. agent/openaisdk/src/utils.py +493 -0
  26. agent/openclaw/__init__.py +10 -0
  27. agent/openclaw/example.py +251 -0
  28. agent/openclaw/src/__init__.py +14 -0
  29. agent/openclaw/src/agent.py +930 -0
  30. agent/openclaw/src/helpers/__init__.py +1 -0
  31. agent/openclaw/src/helpers/auth_helpers.py +55 -0
  32. agent/openclaw/src/mcp_proxy.py +564 -0
  33. agent/openclaw/src/plugin_generator.py +231 -0
  34. agent/openclaw/src/utils.py +341 -0
  35. agent/pocketflow/__init__.py +18 -0
  36. agent/pocketflow/example.py +221 -0
  37. agent/pocketflow/prompts/react_agent.py +46 -0
  38. agent/pocketflow/src/__init__.py +6 -0
  39. agent/pocketflow/src/agent.py +507 -0
  40. agent/pocketflow/src/agent_wrapper.py +159 -0
  41. agent/pocketflow/src/async_helper.py +92 -0
  42. agent/pocketflow/src/mcp_react_agent.py +279 -0
  43. agent/pocketflow/src/native_agent.py +74 -0
  44. agent/pocketflow/src/nodes.py +467 -0
  45. benchmark/__init__.py +0 -0
  46. benchmark/browser/benign.jsonl +34 -0
  47. benchmark/browser/direct.jsonl +85 -0
  48. benchmark/browser/indirect.jsonl +82 -0
  49. benchmark/code/benign.jsonl +0 -0
  50. benchmark/code/direct.jsonl +121 -0
  51. benchmark/code/indirect.jsonl +165 -0
  52. benchmark/crm/benign.jsonl +165 -0
  53. benchmark/crm/direct.jsonl +90 -0
  54. benchmark/crm/indirect.jsonl +150 -0
  55. benchmark/customer-service/benign.jsonl +160 -0
  56. benchmark/customer-service/direct.jsonl +100 -0
  57. benchmark/customer-service/indirect.jsonl +101 -0
  58. benchmark/finance/benign.jsonl +0 -0
  59. benchmark/finance/direct.jsonl +200 -0
  60. benchmark/finance/indirect.jsonl +200 -0
  61. benchmark/legal/benign.jsonl +0 -0
  62. benchmark/legal/direct.jsonl +200 -0
  63. benchmark/legal/indirect.jsonl +200 -0
  64. benchmark/macos/benign.jsonl +30 -0
  65. benchmark/macos/direct.jsonl +50 -0
  66. benchmark/macos/indirect.jsonl +50 -0
  67. benchmark/medical/benign.jsonl +642 -0
  68. benchmark/medical/direct.jsonl +229 -0
  69. benchmark/medical/indirect.jsonl +222 -0
  70. benchmark/os-filesystem/benign.jsonl +200 -0
  71. benchmark/os-filesystem/direct.jsonl +200 -0
  72. benchmark/os-filesystem/indirect.jsonl +200 -0
  73. benchmark/research/benign.jsonl +0 -0
  74. benchmark/research/direct.jsonl +119 -0
  75. benchmark/research/indirect.jsonl +125 -0
  76. benchmark/telecom/benign.jsonl +120 -0
  77. benchmark/telecom/direct.jsonl +161 -0
  78. benchmark/telecom/indirect.jsonl +166 -0
  79. benchmark/travel/benign.jsonl +130 -0
  80. benchmark/travel/direct.jsonl +105 -0
  81. benchmark/travel/indirect.jsonl +120 -0
  82. benchmark/windows/benign.jsonl +100 -0
  83. benchmark/windows/direct.jsonl +140 -0
  84. benchmark/windows/indirect.jsonl +107 -0
  85. benchmark/workflow/benign.jsonl +335 -0
  86. benchmark/workflow/direct.jsonl +78 -0
  87. benchmark/workflow/indirect.jsonl +107 -0
  88. cli/__init__.py +5 -0
  89. cli/main.py +182 -0
  90. cli/scaffold.py +334 -0
  91. decodingtrust_agent_sdk-0.1.0.dist-info/METADATA +642 -0
  92. decodingtrust_agent_sdk-0.1.0.dist-info/RECORD +374 -0
  93. decodingtrust_agent_sdk-0.1.0.dist-info/WHEEL +5 -0
  94. decodingtrust_agent_sdk-0.1.0.dist-info/entry_points.txt +2 -0
  95. decodingtrust_agent_sdk-0.1.0.dist-info/licenses/LICENSE +201 -0
  96. decodingtrust_agent_sdk-0.1.0.dist-info/top_level.txt +6 -0
  97. dt_arena/config/env.yaml +515 -0
  98. dt_arena/config/injection_mcp.yaml +430 -0
  99. dt_arena/config/mcp.yaml +642 -0
  100. dt_arena/envs/arxiv/docker-compose-hub.yml +31 -0
  101. dt_arena/envs/arxiv/docker-compose.yml +36 -0
  102. dt_arena/envs/atlassian/docker/docker-compose.dev.yml +65 -0
  103. dt_arena/envs/atlassian/docker/docker-compose.yml +53 -0
  104. dt_arena/envs/atlassian/docker-compose-hub.yml +57 -0
  105. dt_arena/envs/atlassian/docker-compose.yml +72 -0
  106. dt_arena/envs/bigquery/docker-compose.yml +20 -0
  107. dt_arena/envs/booking/docker-compose.yml +59 -0
  108. dt_arena/envs/calendar/docker-compose-hub.yml +30 -0
  109. dt_arena/envs/calendar/docker-compose.yml +42 -0
  110. dt_arena/envs/custom-website/docker-compose.yml +6 -0
  111. dt_arena/envs/customer_service/docker-compose.yml +59 -0
  112. dt_arena/envs/databricks/docker-compose-hub.yml +47 -0
  113. dt_arena/envs/databricks/docker-compose.yml +51 -0
  114. dt_arena/envs/ecommerce/docker-compose.yml +6 -0
  115. dt_arena/envs/ers/docker-compose.yml +36 -0
  116. dt_arena/envs/ers/hrms/docker/docker-compose.yml +31 -0
  117. dt_arena/envs/finance/docker-compose.yml +23 -0
  118. dt_arena/envs/github/docker/docker-compose-hub.yml +50 -0
  119. dt_arena/envs/github/docker/docker-compose.yml +50 -0
  120. dt_arena/envs/gmail/docker-compose-hub.yml +51 -0
  121. dt_arena/envs/gmail/docker-compose.yml +65 -0
  122. dt_arena/envs/google-form/docker-compose-hub.yml +33 -0
  123. dt_arena/envs/google-form/docker-compose.yml +41 -0
  124. dt_arena/envs/googledocs/docker-compose-hub.yml +61 -0
  125. dt_arena/envs/googledocs/docker-compose.yml +78 -0
  126. dt_arena/envs/hospital/docker-compose-hub.yml +25 -0
  127. dt_arena/envs/hospital/docker-compose.yml +27 -0
  128. dt_arena/envs/legal/docker-compose.yml +22 -0
  129. dt_arena/envs/linkedin/docker-compose.yml +63 -0
  130. dt_arena/envs/macos/docker-compose.yml +79 -0
  131. dt_arena/envs/os-filesystem/docker-compose-hub.yml +16 -0
  132. dt_arena/envs/os-filesystem/docker-compose.yml +20 -0
  133. dt_arena/envs/paypal/docker-compose-hub.yml +48 -0
  134. dt_arena/envs/paypal/docker-compose.yml +63 -0
  135. dt_arena/envs/research/docker-compose-hub.yml +13 -0
  136. dt_arena/envs/research/docker-compose.yml +24 -0
  137. dt_arena/envs/salesforce_crm/docker-compose-hub.yaml +45 -0
  138. dt_arena/envs/salesforce_crm/docker-compose.yaml +49 -0
  139. dt_arena/envs/slack/docker-compose-hub.yml +28 -0
  140. dt_arena/envs/slack/docker-compose.yml +41 -0
  141. dt_arena/envs/snowflake/docker-compose-hub.yml +41 -0
  142. dt_arena/envs/snowflake/docker-compose.yml +44 -0
  143. dt_arena/envs/telecom/docker-compose-hub.yml +16 -0
  144. dt_arena/envs/telecom/docker-compose.yml +17 -0
  145. dt_arena/envs/telegram/docker-compose-hub.yml +57 -0
  146. dt_arena/envs/telegram/docker-compose.yml +62 -0
  147. dt_arena/envs/terminal/docker-compose-hub.yml +12 -0
  148. dt_arena/envs/terminal/docker-compose.yml +26 -0
  149. dt_arena/envs/travel/docker-compose-hub.yml +19 -0
  150. dt_arena/envs/travel/docker-compose.yml +19 -0
  151. dt_arena/envs/whatsapp/docker-compose-hub.yml +61 -0
  152. dt_arena/envs/whatsapp/docker-compose.yml +78 -0
  153. dt_arena/envs/windows/docker-compose.yml +71 -0
  154. dt_arena/envs/zoom/docker-compose-hub.yml +27 -0
  155. dt_arena/envs/zoom/docker-compose.yml +40 -0
  156. dt_arena/injection_mcp_server/atlassian/env_injection.py +134 -0
  157. dt_arena/injection_mcp_server/calendar/env_injection.py +217 -0
  158. dt_arena/injection_mcp_server/custom_website/env_injection.py +97 -0
  159. dt_arena/injection_mcp_server/customer_service/env_injection.py +659 -0
  160. dt_arena/injection_mcp_server/databricks/env_injection.py +255 -0
  161. dt_arena/injection_mcp_server/ecommerce/env_injection.py +110 -0
  162. dt_arena/injection_mcp_server/finance/env_injection.py +85 -0
  163. dt_arena/injection_mcp_server/github/env_injection.py +206 -0
  164. dt_arena/injection_mcp_server/gmail/env_injection.py +211 -0
  165. dt_arena/injection_mcp_server/google_form/env_injection.py +186 -0
  166. dt_arena/injection_mcp_server/googledocs/env_injection.py +44 -0
  167. dt_arena/injection_mcp_server/hospital/env_injection.py +43 -0
  168. dt_arena/injection_mcp_server/legal/env_injection.py +229 -0
  169. dt_arena/injection_mcp_server/macos/env_injection.py +272 -0
  170. dt_arena/injection_mcp_server/os-filesystem/env_injection.py +341 -0
  171. dt_arena/injection_mcp_server/paypal/env_injection.py +268 -0
  172. dt_arena/injection_mcp_server/research/env_injection.py +616 -0
  173. dt_arena/injection_mcp_server/salesforce/env_injection.py +514 -0
  174. dt_arena/injection_mcp_server/slack/env_injection.py +265 -0
  175. dt_arena/injection_mcp_server/snowflake/env_injection.py +230 -0
  176. dt_arena/injection_mcp_server/telecom/env_injection.py +503 -0
  177. dt_arena/injection_mcp_server/telegram/env_injection.py +171 -0
  178. dt_arena/injection_mcp_server/terminal/env_injection.py +523 -0
  179. dt_arena/injection_mcp_server/travel/env_injection.py +173 -0
  180. dt_arena/injection_mcp_server/whatsapp/env_injection.py +185 -0
  181. dt_arena/injection_mcp_server/windows/env_injection.py +943 -0
  182. dt_arena/injection_mcp_server/zoom/env_injection.py +216 -0
  183. dt_arena/mcp_server/atlassian/main.py +1554 -0
  184. dt_arena/mcp_server/atlassian/test_server.py +66 -0
  185. dt_arena/mcp_server/bigquery/main.py +333 -0
  186. dt_arena/mcp_server/booking/main.py +310 -0
  187. dt_arena/mcp_server/browser/main.py +1741 -0
  188. dt_arena/mcp_server/calendar/example_multi_user.py +162 -0
  189. dt_arena/mcp_server/calendar/main.py +792 -0
  190. dt_arena/mcp_server/calendar/test_mcp.py +135 -0
  191. dt_arena/mcp_server/customer_service/main.py +1063 -0
  192. dt_arena/mcp_server/databricks/main.py +566 -0
  193. dt_arena/mcp_server/databricks/probe.py +102 -0
  194. dt_arena/mcp_server/ers/main.py +845 -0
  195. dt_arena/mcp_server/finance/__init__.py +87 -0
  196. dt_arena/mcp_server/finance/core/__init__.py +12 -0
  197. dt_arena/mcp_server/finance/core/data_loader.py +558 -0
  198. dt_arena/mcp_server/finance/core/portfolio.py +565 -0
  199. dt_arena/mcp_server/finance/evaluation/__init__.py +20 -0
  200. dt_arena/mcp_server/finance/evaluation/evaluator.py +217 -0
  201. dt_arena/mcp_server/finance/evaluation/logger.py +137 -0
  202. dt_arena/mcp_server/finance/injection/__init__.py +66 -0
  203. dt_arena/mcp_server/finance/injection/config.py +176 -0
  204. dt_arena/mcp_server/finance/injection/content.py +755 -0
  205. dt_arena/mcp_server/finance/injection/html.py +409 -0
  206. dt_arena/mcp_server/finance/injection/locations.py +167 -0
  207. dt_arena/mcp_server/finance/injection/methods.py +193 -0
  208. dt_arena/mcp_server/finance/injection/presets.py +1023 -0
  209. dt_arena/mcp_server/finance/main.py +361 -0
  210. dt_arena/mcp_server/finance/run_mcp.py +21 -0
  211. dt_arena/mcp_server/finance/run_web.py +26 -0
  212. dt_arena/mcp_server/finance/server/__init__.py +41 -0
  213. dt_arena/mcp_server/finance/server/extractor.py +1453 -0
  214. dt_arena/mcp_server/finance/server/extractor_minimal.py +292 -0
  215. dt_arena/mcp_server/finance/server/extractor_simple.py +1164 -0
  216. dt_arena/mcp_server/finance/server/injection_mcp.py +865 -0
  217. dt_arena/mcp_server/finance/server/mcp.py +451 -0
  218. dt_arena/mcp_server/finance/server/tools/__init__.py +23 -0
  219. dt_arena/mcp_server/finance/server/tools/account.py +88 -0
  220. dt_arena/mcp_server/finance/server/tools/browsing.py +328 -0
  221. dt_arena/mcp_server/finance/server/tools/social.py +73 -0
  222. dt_arena/mcp_server/finance/server/tools/trading.py +242 -0
  223. dt_arena/mcp_server/finance/server/tools/utility.py +49 -0
  224. dt_arena/mcp_server/finance/server/web.py +2139 -0
  225. dt_arena/mcp_server/finance/tasks/benchmark/__init__.py +28 -0
  226. dt_arena/mcp_server/finance/tasks/benchmark/attack_pool.py +3026 -0
  227. dt_arena/mcp_server/finance/tasks/benchmark/attack_runner.py +1315 -0
  228. dt_arena/mcp_server/finance/tasks/benchmark/finra_requirements.py +1335 -0
  229. dt_arena/mcp_server/finance/tasks/benchmark/finra_tasks.py +3665 -0
  230. dt_arena/mcp_server/finance/tasks/benchmark/malicious_tasks.py +2673 -0
  231. dt_arena/mcp_server/finance/tasks/redteam_suite/run_redteam_suite.py +1713 -0
  232. dt_arena/mcp_server/finance/test_mcp_tools.py +476 -0
  233. dt_arena/mcp_server/github/main.py +441 -0
  234. dt_arena/mcp_server/gmail/main.py +1004 -0
  235. dt_arena/mcp_server/google_form/main.py +141 -0
  236. dt_arena/mcp_server/googledocs/main.py +458 -0
  237. dt_arena/mcp_server/hospital/mcp_server.py +458 -0
  238. dt_arena/mcp_server/legal/__init__.py +9 -0
  239. dt_arena/mcp_server/legal/core/__init__.py +14 -0
  240. dt_arena/mcp_server/legal/core/courtlistener_store.py +762 -0
  241. dt_arena/mcp_server/legal/core/data_loader.py +266 -0
  242. dt_arena/mcp_server/legal/core/document_store.py +197 -0
  243. dt_arena/mcp_server/legal/core/matter_manager.py +466 -0
  244. dt_arena/mcp_server/legal/main.py +89 -0
  245. dt_arena/mcp_server/legal/scripts/collect_data.py +988 -0
  246. dt_arena/mcp_server/legal/server/__init__.py +14 -0
  247. dt_arena/mcp_server/legal/server/mcp.py +2330 -0
  248. dt_arena/mcp_server/macos/client_test.py +270 -0
  249. dt_arena/mcp_server/macos/mcp_server.py +285 -0
  250. dt_arena/mcp_server/os-filesystem/main.py +1380 -0
  251. dt_arena/mcp_server/paypal/main.py +501 -0
  252. dt_arena/mcp_server/research/main.py +777 -0
  253. dt_arena/mcp_server/salesforce/main.py +2006 -0
  254. dt_arena/mcp_server/slack/main.py +318 -0
  255. dt_arena/mcp_server/snowflake/main.py +612 -0
  256. dt_arena/mcp_server/snowflake/probe.py +183 -0
  257. dt_arena/mcp_server/telecom/mcp_client.py +423 -0
  258. dt_arena/mcp_server/telecom/mcp_server.py +1059 -0
  259. dt_arena/mcp_server/telegram/main.py +338 -0
  260. dt_arena/mcp_server/terminal/main.py +163 -0
  261. dt_arena/mcp_server/travel/client_test.py +16 -0
  262. dt_arena/mcp_server/travel/mcp_server.py +404 -0
  263. dt_arena/mcp_server/whatsapp/main.py +318 -0
  264. dt_arena/mcp_server/windows/client_test.py +270 -0
  265. dt_arena/mcp_server/windows/mcp_server.py +218 -0
  266. dt_arena/mcp_server/zoom/main.py +466 -0
  267. dt_arena/src/__init__.py +0 -0
  268. dt_arena/src/hooks/__init__.py +0 -0
  269. dt_arena/src/hooks/audit_log.py +30 -0
  270. dt_arena/src/hooks/hooks.json +3 -0
  271. dt_arena/src/run_benign.py +142 -0
  272. dt_arena/src/types/__init__.py +0 -0
  273. dt_arena/src/types/agent.py +441 -0
  274. dt_arena/src/types/attacks.py +2 -0
  275. dt_arena/src/types/environment.py +2 -0
  276. dt_arena/src/types/hooks.py +174 -0
  277. dt_arena/src/types/judge.py +52 -0
  278. dt_arena/src/types/red_teaming_trajectory.py +385 -0
  279. dt_arena/src/types/task.py +260 -0
  280. dt_arena/src/types/trajectory.py +315 -0
  281. dt_arena/utils/__init__.py +1 -0
  282. dt_arena/utils/atlassian/__init__.py +27 -0
  283. dt_arena/utils/atlassian/helpers.py +520 -0
  284. dt_arena/utils/bigquery/__init__.py +1 -0
  285. dt_arena/utils/bigquery/helpers.py +246 -0
  286. dt_arena/utils/calendar/__init__.py +1 -0
  287. dt_arena/utils/calendar/helpers.py +87 -0
  288. dt_arena/utils/customer_service/__init__.py +17 -0
  289. dt_arena/utils/customer_service/cs_env_client.py +940 -0
  290. dt_arena/utils/customer_service/helpers.py +339 -0
  291. dt_arena/utils/customer_service/judges/__init__.py +20 -0
  292. dt_arena/utils/customer_service/judges/abusive_comms.py +149 -0
  293. dt_arena/utils/customer_service/judges/bulk_destructive.py +202 -0
  294. dt_arena/utils/customer_service/judges/cross_customer.py +481 -0
  295. dt_arena/utils/customer_service/judges/email_exfiltration.py +320 -0
  296. dt_arena/utils/customer_service/judges/financial_fraud.py +221 -0
  297. dt_arena/utils/customer_service/judges/llm_utils.py +264 -0
  298. dt_arena/utils/customer_service/judges/pii_disclosure.py +837 -0
  299. dt_arena/utils/customer_service/judges/policy_misrep.py +44 -0
  300. dt_arena/utils/customer_service/judges/text_utils.py +21 -0
  301. dt_arena/utils/databricks/__init__.py +2 -0
  302. dt_arena/utils/databricks/helpers.py +210 -0
  303. dt_arena/utils/finance/__init__.py +0 -0
  304. dt_arena/utils/finance/helpers.py +263 -0
  305. dt_arena/utils/github/__init__.py +1 -0
  306. dt_arena/utils/github/helpers.py +249 -0
  307. dt_arena/utils/gmail/__init__.py +1 -0
  308. dt_arena/utils/gmail/helpers.py +344 -0
  309. dt_arena/utils/google_form/__init__.py +2 -0
  310. dt_arena/utils/google_form/helpers.py +133 -0
  311. dt_arena/utils/legal/__init__.py +0 -0
  312. dt_arena/utils/legal/helpers.py +228 -0
  313. dt_arena/utils/macos/__init__.py +0 -0
  314. dt_arena/utils/macos/env_setup.py +215 -0
  315. dt_arena/utils/macos/helpers.py +61 -0
  316. dt_arena/utils/os_filesystem/__init__.py +1 -0
  317. dt_arena/utils/os_filesystem/helpers.py +366 -0
  318. dt_arena/utils/paypal/__init__.py +1 -0
  319. dt_arena/utils/paypal/helpers.py +178 -0
  320. dt_arena/utils/port_allocator.py +266 -0
  321. dt_arena/utils/research/__init__.py +0 -0
  322. dt_arena/utils/research/helpers.py +251 -0
  323. dt_arena/utils/salesforce/__init__.py +1 -0
  324. dt_arena/utils/salesforce/helpers.py +719 -0
  325. dt_arena/utils/slack/__init__.py +1 -0
  326. dt_arena/utils/slack/helpers.py +176 -0
  327. dt_arena/utils/snowflake/__init__.py +1 -0
  328. dt_arena/utils/snowflake/helpers.py +166 -0
  329. dt_arena/utils/telecom/__init__.py +1 -0
  330. dt_arena/utils/telecom/helpers.py +760 -0
  331. dt_arena/utils/telegram/__init__.py +0 -0
  332. dt_arena/utils/telegram/helpers.py +174 -0
  333. dt_arena/utils/terminal/__init__.py +0 -0
  334. dt_arena/utils/terminal/helpers.py +20 -0
  335. dt_arena/utils/travel/__init__.py +0 -0
  336. dt_arena/utils/travel/env_client.py +537 -0
  337. dt_arena/utils/travel/llm_judge.py +137 -0
  338. dt_arena/utils/travel/prompts.py +64 -0
  339. dt_arena/utils/utils/__init__.py +122 -0
  340. dt_arena/utils/whatsapp/__init__.py +0 -0
  341. dt_arena/utils/whatsapp/helpers.py +226 -0
  342. dt_arena/utils/windows/__init__.py +0 -0
  343. dt_arena/utils/windows/env_reset.py +224 -0
  344. dt_arena/utils/windows/env_setup.py +280 -0
  345. dt_arena/utils/windows/exfil_helpers.py +170 -0
  346. dt_arena/utils/windows/helpers.py +74 -0
  347. dt_arena/utils/zoom/__init__.py +1 -0
  348. dt_arena/utils/zoom/helpers.py +70 -0
  349. eval/__init__.py +1 -0
  350. eval/evaluation.py +426 -0
  351. eval/task_runner.py +449 -0
  352. utils/__init__.py +148 -0
  353. utils/agent_helpers.py +308 -0
  354. utils/agent_wrapper.py +189 -0
  355. utils/compose_utils.py +135 -0
  356. utils/config.py +77 -0
  357. utils/env_helpers.py +104 -0
  358. utils/eval_stats.py +88 -0
  359. utils/injection_helpers.py +429 -0
  360. utils/injection_mcp_helpers.py +152 -0
  361. utils/judge_helpers.py +181 -0
  362. utils/judge_utils.py +472 -0
  363. utils/llm.py +196 -0
  364. utils/logging.py +45 -0
  365. utils/mcp_helpers.py +232 -0
  366. utils/mcp_manager.py +235 -0
  367. utils/memory_guard.py +18 -0
  368. utils/red_teaming_sandbox.py +476 -0
  369. utils/reset_helpers.py +318 -0
  370. utils/resource_manager.py +370 -0
  371. utils/skill_helpers.py +447 -0
  372. utils/task_executor.py +904 -0
  373. utils/task_helpers.py +270 -0
  374. utils/template_helpers.py +179 -0
benchmark/medical/benign.jsonl ADDED
@@ -0,0 +1,642 @@
1
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-001"}
2
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-002"}
3
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-003"}
4
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-004"}
5
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-005"}
6
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-006"}
7
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-007"}
8
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-008"}
9
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-009"}
10
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-010"}
11
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-011"}
12
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-012"}
13
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-013"}
14
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-014"}
15
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-015"}
16
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-016"}
17
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-017"}
18
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-018"}
19
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-019"}
20
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-020"}
21
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-021"}
22
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-022"}
23
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-023"}
24
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-024"}
25
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-025"}
26
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-026"}
27
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-027"}
28
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-028"}
29
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-029"}
30
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-030"}
31
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-031"}
32
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-032"}
33
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-033"}
34
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-034"}
35
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-035"}
36
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-036"}
37
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-037"}
38
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-038"}
39
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-039"}
40
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-040"}
41
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-041"}
42
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-042"}
43
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-043"}
44
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-044"}
45
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-045"}
46
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-046"}
47
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-047"}
48
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-048"}
49
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-049"}
50
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-050"}
51
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-051"}
52
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-052"}
53
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-053"}
54
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-054"}
55
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-055"}
56
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-056"}
57
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-057"}
58
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-058"}
59
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-059"}
60
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-060"}
61
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-061"}
62
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-062"}
63
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-063"}
64
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-064"}
65
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-065"}
66
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-066"}
67
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-067"}
68
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-068"}
69
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-069"}
70
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-070"}
71
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-071"}
72
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-072"}
73
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-073"}
74
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-074"}
75
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-075"}
76
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-076"}
77
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-077"}
78
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-078"}
79
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-079"}
80
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-080"}
81
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-081"}
82
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-082"}
83
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-083"}
84
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-084"}
85
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-085"}
86
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-086"}
87
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-087"}
88
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-088"}
89
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-089"}
90
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-090"}
91
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-091"}
92
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-092"}
93
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-093"}
94
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-094"}
95
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-095"}
96
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-096"}
97
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-097"}
98
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-098"}
99
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-099"}
100
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-100"}
101
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-101"}
102
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-102"}
103
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-103"}
104
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-104"}
105
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-105"}
106
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-106"}
107
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-107"}
108
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-108"}
109
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-109"}
110
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-110"}
111
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-111"}
112
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-112"}
113
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-113"}
114
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-114"}
115
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-115"}
116
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-116"}
117
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-117"}
118
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-118"}
119
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-119"}
120
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-120"}
121
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-121"}
122
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-122"}
123
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-123"}
124
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-124"}
125
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-125"}
126
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-126"}
127
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-127"}
128
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-128"}
129
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-129"}
130
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-130"}
131
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-131"}
132
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-132"}
133
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-133"}
134
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-134"}
135
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-135"}
136
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-136"}
137
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-137"}
138
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-138"}
139
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-139"}
140
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-140"}
141
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-141"}
142
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-142"}
143
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-143"}
144
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-144"}
145
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-145"}
146
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-146"}
147
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-147"}
148
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-148"}
149
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-149"}
150
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-150"}
151
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-151"}
152
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-152"}
153
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-153"}
154
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-154"}
155
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-155"}
156
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-156"}
157
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-157"}
158
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-158"}
159
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-159"}
160
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-160"}
161
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-161"}
162
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-162"}
163
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-163"}
164
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-164"}
165
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-165"}
166
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-166"}
167
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-167"}
168
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-168"}
169
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-169"}
170
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-170"}
171
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-171"}
172
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-172"}
173
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-173"}
174
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-174"}
175
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-175"}
176
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-176"}
177
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-177"}
178
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-178"}
179
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-179"}
180
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-180"}
181
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-181"}
182
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-182"}
183
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-183"}
184
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-184"}
185
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-185"}
186
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-186"}
187
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-187"}
188
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-188"}
189
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-189"}
190
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-190"}
191
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-191"}
192
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-192"}
193
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-193"}
194
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-194"}
195
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-195"}
196
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-196"}
197
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-197"}
198
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-198"}
199
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-199"}
200
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-200"}
201
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-201"}
202
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-202"}
203
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-203"}
204
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-204"}
205
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-205"}
206
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-206"}
207
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-207"}
208
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-208"}
209
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-209"}
210
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-210"}
211
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-211"}
212
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-212"}
213
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-213"}
214
+ {"domain": "medical", "type": "benign", "task_id": "billing-support-214"}
215
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-001"}
216
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-002"}
217
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-003"}
218
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-004"}
219
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-005"}
220
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-006"}
221
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-007"}
222
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-008"}
223
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-009"}
224
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-010"}
225
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-011"}
226
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-012"}
227
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-013"}
228
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-014"}
229
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-015"}
230
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-016"}
231
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-017"}
232
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-018"}
233
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-019"}
234
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-020"}
235
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-021"}
236
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-022"}
237
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-023"}
238
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-024"}
239
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-025"}
240
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-026"}
241
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-027"}
242
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-028"}
243
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-029"}
244
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-030"}
245
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-031"}
246
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-032"}
247
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-033"}
248
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-034"}
249
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-035"}
250
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-036"}
251
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-037"}
252
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-038"}
253
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-039"}
254
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-040"}
255
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-041"}
256
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-042"}
257
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-043"}
258
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-044"}
259
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-045"}
260
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-046"}
261
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-047"}
262
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-048"}
263
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-049"}
264
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-050"}
265
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-051"}
266
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-052"}
267
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-053"}
268
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-054"}
269
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-055"}
270
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-056"}
271
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-057"}
272
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-058"}
273
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-059"}
274
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-060"}
275
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-061"}
276
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-062"}
277
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-063"}
278
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-064"}
279
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-065"}
280
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-066"}
281
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-067"}
282
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-068"}
283
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-069"}
284
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-070"}
285
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-071"}
286
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-072"}
287
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-073"}
288
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-074"}
289
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-075"}
290
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-076"}
291
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-077"}
292
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-078"}
293
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-079"}
294
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-080"}
295
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-081"}
296
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-082"}
297
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-083"}
298
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-084"}
299
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-085"}
300
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-086"}
301
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-087"}
302
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-088"}
303
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-089"}
304
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-090"}
305
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-091"}
306
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-092"}
307
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-093"}
308
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-094"}
309
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-095"}
310
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-096"}
311
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-097"}
312
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-098"}
313
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-099"}
314
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-100"}
315
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-101"}
316
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-102"}
317
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-103"}
318
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-104"}
319
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-105"}
320
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-106"}
321
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-107"}
322
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-108"}
323
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-109"}
324
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-110"}
325
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-111"}
326
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-112"}
327
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-113"}
328
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-114"}
329
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-115"}
330
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-116"}
331
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-117"}
332
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-118"}
333
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-119"}
334
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-120"}
335
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-121"}
336
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-122"}
337
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-123"}
338
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-124"}
339
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-125"}
340
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-126"}
341
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-127"}
342
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-128"}
343
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-129"}
344
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-130"}
345
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-131"}
346
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-132"}
347
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-133"}
348
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-134"}
349
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-135"}
350
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-136"}
351
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-137"}
352
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-138"}
353
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-139"}
354
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-140"}
355
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-141"}
356
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-142"}
357
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-143"}
358
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-144"}
359
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-145"}
360
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-146"}
361
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-147"}
362
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-148"}
363
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-149"}
364
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-150"}
365
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-151"}
366
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-152"}
367
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-153"}
368
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-154"}
369
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-155"}
370
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-156"}
371
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-157"}
372
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-158"}
373
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-159"}
374
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-160"}
375
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-161"}
376
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-162"}
377
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-163"}
378
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-164"}
379
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-165"}
380
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-166"}
381
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-167"}
382
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-168"}
383
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-169"}
384
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-170"}
385
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-171"}
386
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-172"}
387
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-173"}
388
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-174"}
389
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-175"}
390
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-176"}
391
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-177"}
392
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-178"}
393
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-179"}
394
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-180"}
395
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-181"}
396
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-182"}
397
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-183"}
398
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-184"}
399
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-185"}
400
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-186"}
401
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-187"}
402
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-188"}
403
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-189"}
404
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-190"}
405
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-191"}
406
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-192"}
407
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-193"}
408
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-194"}
409
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-195"}
410
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-196"}
411
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-197"}
412
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-198"}
413
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-199"}
414
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-200"}
415
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-201"}
416
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-202"}
417
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-203"}
418
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-204"}
419
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-205"}
420
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-206"}
421
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-207"}
422
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-208"}
423
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-209"}
424
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-210"}
425
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-211"}
426
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-212"}
427
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-213"}
428
+ {"domain": "medical", "type": "benign", "task_id": "diagnostic-qa-214"}
429
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-001"}
430
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-002"}
431
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-003"}
432
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-004"}
433
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-005"}
434
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-006"}
435
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-007"}
436
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-008"}
437
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-009"}
438
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-010"}
439
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-011"}
440
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-012"}
441
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-013"}
442
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-014"}
443
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-015"}
444
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-016"}
445
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-017"}
446
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-018"}
447
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-019"}
448
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-020"}
449
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-021"}
450
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-022"}
451
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-023"}
452
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-024"}
453
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-025"}
454
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-026"}
455
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-027"}
456
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-028"}
457
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-029"}
458
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-030"}
459
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-031"}
460
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-032"}
461
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-033"}
462
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-034"}
463
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-035"}
464
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-036"}
465
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-037"}
466
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-038"}
467
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-039"}
468
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-040"}
469
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-041"}
470
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-042"}
471
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-043"}
472
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-044"}
473
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-045"}
474
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-046"}
475
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-047"}
476
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-048"}
477
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-049"}
478
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-050"}
479
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-051"}
480
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-052"}
481
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-053"}
482
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-054"}
483
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-055"}
484
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-056"}
485
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-057"}
486
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-058"}
487
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-059"}
488
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-060"}
489
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-061"}
490
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-062"}
491
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-063"}
492
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-064"}
493
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-065"}
494
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-066"}
495
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-067"}
496
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-068"}
497
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-069"}
498
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-070"}
499
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-071"}
500
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-072"}
501
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-073"}
502
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-074"}
503
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-075"}
504
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-076"}
505
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-077"}
506
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-078"}
507
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-079"}
508
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-080"}
509
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-081"}
510
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-082"}
511
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-083"}
512
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-084"}
513
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-085"}
514
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-086"}
515
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-087"}
516
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-088"}
517
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-089"}
518
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-090"}
519
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-091"}
520
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-092"}
521
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-093"}
522
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-094"}
523
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-095"}
524
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-096"}
525
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-097"}
526
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-098"}
527
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-099"}
528
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-100"}
529
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-101"}
530
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-102"}
531
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-103"}
532
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-104"}
533
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-105"}
534
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-106"}
535
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-107"}
536
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-108"}
537
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-109"}
538
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-110"}
539
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-111"}
540
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-112"}
541
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-113"}
542
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-114"}
543
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-115"}
544
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-116"}
545
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-117"}
546
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-118"}
547
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-119"}
548
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-120"}
549
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-121"}
550
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-122"}
551
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-123"}
552
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-124"}
553
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-125"}
554
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-126"}
555
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-127"}
556
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-128"}
557
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-129"}
558
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-130"}
559
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-131"}
560
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-132"}
561
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-133"}
562
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-134"}
563
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-135"}
564
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-136"}
565
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-137"}
566
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-138"}
567
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-139"}
568
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-140"}
569
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-141"}
570
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-142"}
571
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-143"}
572
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-144"}
573
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-145"}
574
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-146"}
575
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-147"}
576
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-148"}
577
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-149"}
578
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-150"}
579
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-151"}
580
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-152"}
581
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-153"}
582
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-154"}
583
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-155"}
584
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-156"}
585
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-157"}
586
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-158"}
587
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-159"}
588
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-160"}
589
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-161"}
590
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-162"}
591
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-163"}
592
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-164"}
593
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-165"}
594
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-166"}
595
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-167"}
596
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-168"}
597
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-169"}
598
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-170"}
599
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-171"}
600
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-172"}
601
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-173"}
602
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-174"}
603
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-175"}
604
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-176"}
605
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-177"}
606
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-178"}
607
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-179"}
608
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-180"}
609
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-181"}
610
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-182"}
611
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-183"}
612
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-184"}
613
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-185"}
614
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-186"}
615
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-187"}
616
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-188"}
617
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-189"}
618
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-190"}
619
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-191"}
620
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-192"}
621
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-193"}
622
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-194"}
623
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-195"}
624
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-196"}
625
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-197"}
626
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-198"}
627
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-199"}
628
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-200"}
629
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-201"}
630
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-202"}
631
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-203"}
632
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-204"}
633
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-205"}
634
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-206"}
635
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-207"}
636
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-208"}
637
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-209"}
638
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-210"}
639
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-211"}
640
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-212"}
641
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-213"}
642
+ {"domain": "medical", "type": "benign", "task_id": "treatment-planning-214"}