InstructEval - Problem Solving: leaderboard
DeCLaRe Lab's 2023 instruction-tuned suite, averaging MMLU, BBH, DROP, CRASS and HumanEval under one harness. A historical board: the roster is the Alpaca / Vicuna / Flan generation, so it anchors the low end of the corpus rather than tracking the frontier.
Metric: Average (%, MMLU/BBH/DROP/CRASS/HumanEval). Source: declare-lab.github.io. Saturation forecast: Estimated already saturated. 54 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | ChatGPT | 64.5 | #900 |
| 2 | alpaca-7B | 32.4 | #1611 |
| 3 | starcoder | 27.5 | #1577 |
| 4 | pythia-12B | 23.9 | #1621 |
| 5 | falcon-7B | 23.7 | #1598 |
| 6 | falcon-7B Instruct | 22.2 | #1622 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=instructeval-problem-solving · How It Works · Data refreshed daily, snapshot 2026-10-20.