mkurman |------|--------|
| mmlu | Match | Knowledge (57 subjects) |
| hellaswag | Match | Commonsense reasoning |
| truthfulqa | Model-graded | Truthfulness |
| gsm8k | Match | Math reasoning |
| humaneval | Custom | Code generation |
| ifeval | Model-graded | Instruction following |
| bbq | Model-graded | Bias detection |
| factuality | Model-graded | Factual accuracy |
| translation | Model-graded | Translation quality |
Browse full registry: evals/registry/evals/
10. Production Eval Pipeline Pattern
# CI/CD eval pipeline
def run_eval_suite(model_name, eval_names):
results = {}
for eval_name in eval_names:
cmd = f"oaieval {model_name} {eval_name} --max_samples 200"
result = subprocess.run(cmd, shell=True, capture_output=True)
results[eval_name] = parse_accuracy(result.stdout)
return results
# Regression test
previous = {"mmlu": 0.86, "gsm8k": 0.92, "hellaswag": 0.85}
current = run_eval_suite("my-finetuned-model", ["mmlu", "gsm8k", "hellaswag"])
for name, score in current.items():
if score < previous[name] - 0.02: # 2% regression threshold
alert(f"Regression in {name}: {previous[name]:.2f} → {score:.2f}")
Key Patterns
- Start with templates — most evals don't need custom Python code
- Use model-graded evals for subjective quality (fluency, helpfulness, safety)
- Use match evals for objective metrics (classification, multiple choice)
git lfs fetch --allis required before running community evals- Custom completion functions enable testing non-OpenAI models
- Record paths enable debugging individual failures
- Private evals can test proprietary data without exposing it