InstructEval - Alignment (HHH): leaderboard
The alignment half of InstructEval, scoring the same 2023 roster on harmlessness, helpfulness and honesty. Useful as an early snapshot of how instruction tuning traded capability against refusal behaviour, not as a current safety measure.
Metric: Average (%, harmless/helpful/honest). Source: declare-lab.github.io. Saturation forecast: Around January 2027. 28 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | ChatGPT | 86.6 | #900 |
| 2 | pythia-12B | 45.1 | #1621 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=instructeval-alignment-hhh · How It Works · Data refreshed daily, snapshot 2026-10-20.