What's the most effective safety eval suite for scoring model helpfulness versus harmlessness during release reviews?

There isn’t a single universally “best” suite, but for release-review scoring of helpfulness vs. harmlessness, the most effective setup in practice is usually a composite eval suite, not one benchmark.

Best practical choice

If you want something that works well for release decisions, I’d recommend:

1) Safety-oriented behavior suite

Use a suite that covers:

• refusal quality
• policy adherence
• harmful instruction handling
• jailbreak robustness
• over-refusal / false refusal
• safe completion on benign but sensitive prompts

Good components to include:

• HarmBench: strong for harmful instruction following / refusal behavior
• StrongREJECT or similar jailbreak-style red-teaming sets: good for adversarial robustness
• RealToxicityPrompts or toxicity-focused evals: useful but not sufficient alone
• AdvBench / JailbreakBench-style sets: helpful for adversarial safety stress tests

2) Helpfulness eval suite

Use a separate suite for:

• task completion
• instruction following
• reasoning quality
• summarization / coding / QA
• user satisfaction proxies

Common choices:

• MT-Bench / Arena-style pairwise evals for general chat helpfulness
• MMLU / GPQA / GSM8K / HumanEval depending on your model’s target capabilities
• IFEval for instruction-following reliability

3) Joint “helpful-but-safe” evaluation

For release review, the most important metric is usually not raw helpfulness or raw harmlessness, but:

• helpfulness on benign prompts
• refusal on disallowed prompts
• appropriate safe completion on dual-use / sensitive prompts
• low over-refusal rate

This is where you score models on a matrix like:

• benign helpfulness
• sensitive-but-allowed helpfulness
• clearly disallowed content refusal
• jailbreak resistance
• calibration/uncertainty behavior

If you want one “best” suite-like answer

A strong default is:

• Helpfulness: MT-Bench + IFEval + a task-specific benchmark
• Harmlessness: HarmBench + jailbreak/adversarial set + toxicity set
• Release gate: a weighted rubric measuring:
  • benign usefulness
  • disallowed content refusal
  • safe completion
  • over-refusal
  • jailbreak robustness

Recommendation by use case

If you need a single number for release review

Build a weighted composite score from:

• helpfulness pass rate
• harmful compliance failure rate
• over-refusal rate
• jailbreak success rate against the model
• safe-completion rate on ambiguous prompts

If you need the most defensible governance process

Use:

• automated evals
• human red-teaming
• pairwise model comparisons
• policy-based labeling of failures

Bottom line

The most effective approach is not one suite, but a multi-axis eval stack:

• HarmBench-style safety
• MT-Bench/IFEval-style helpfulness
• adversarial jailbreak tests
• over-refusal checks
• human review for edge cases

If you want, I can propose a concrete release-review scoring rubric with weights and pass/fail thresholds.