College biology isn't medicine

#4
Files changed (1) hide show
  1. src/backend/run_eval_suite.py +1 -1
src/backend/run_eval_suite.py CHANGED
@@ -72,7 +72,7 @@ def run_evaluation(eval_request: EvalRequest, task_names, num_fewshot, batch_siz
72
  print(
73
  "WARNING: --limit SHOULD ONLY BE USED FOR TESTING. REAL METRICS SHOULD NOT BE COMPUTED USING LIMIT."
74
  )
75
- task_names = ["medmcqa", "medqa_4options", "mmlu_anatomy", "mmlu_clinical_knowledge", "mmlu_college_biology", "mmlu_college_medicine", "mmlu_medical_genetics", "mmlu_professional_medicine", "pubmedqa"]
76
 
77
 
78
  print(f"Selected Tasks: {task_names}")
 
72
  print(
73
  "WARNING: --limit SHOULD ONLY BE USED FOR TESTING. REAL METRICS SHOULD NOT BE COMPUTED USING LIMIT."
74
  )
75
+ task_names = ["medmcqa", "medqa_4options", "mmlu_anatomy", "mmlu_clinical_knowledge", "mmlu_college_medicine", "mmlu_medical_genetics", "mmlu_professional_medicine", "pubmedqa"]
76
 
77
 
78
  print(f"Selected Tasks: {task_names}")