Wmdp Eval

wmdp-eval

The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning — Li et al. (2024) (arXiv:2403.03218, 2024)

What this evaluates

Evaluates large language models' knowledge of hazardous topics in biosecurity, cybersecurity, and chemical security, as well as their general knowledge and fluency. It serves as a proxy for measuring dual-use risk and benchmarking unlearning methods.

Datasets

• WMDP — total 3668; splits: test (3668)

Metrics

• WMDP (primary) — range: percent
  • Percentage of correctly answered multiple-choice questions using top-1 logit selection. Lower scores indicate less hazardous knowledge.
• MMLU — range: percent
  • Percentage of correctly answered multiple-choice questions across college and high school subjects.
• MT-Bench — range: other
  • Single-evaluation mode scored by GPT-4-0613 as an automated judge.

Input / output format

Input: Zero-shot multiple-choice question prompt with four options (A, B, C, D) and a trailing 'Answer:' token.

Output: Model generates the top logit choice (A, B, C, or D).

Scoring recipe

def compute_accuracy(predictions, gold):
    correct = sum(1 for p, g in zip(predictions, gold) if p == g)
    return (correct / len(gold)) * 100

Common pitfalls

• WMDP scores are lower-is-better (indicated by ↓ in tables), unlike standard benchmarks where higher is better.
• MT-Bench evaluation relies on GPT-4 as an automated judge rather than exact-match or human grading.
• The benchmark uses a strict zero-shot format with a fixed prompt template; deviating from the template may alter model behavior.

Evidence (verbatim from paper)

We use the default lm-evaluation-harness zero-shot format for MMLU and WMDP, taking the top logit between A, B, C, and D as the answer choice. ... Table 2: Comparing base models and unlearning methods on question-answer evaluation (WMDP, MMLU) and fluency (MT-Bench). All WMDP and MMLU scores are percentage points.

Citation

@misc{li2024wmdp,
  title={The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning},
  author={Li et al. (2024)},
  year={2024},
  note={arXiv:2403.03218}
}

• arXiv: 2403.03218