HELM Classic - RAFT: leaderboard
The "RAFT" subtask of HELM Classic. A scenario from Stanford HELM Classic, the holistic evaluation suite scoring language models across accuracy, robustness and calibration on a broad set of tasks.
Metric: Exact Match (%). Source: crfm.stanford.edu. Saturation forecast: Estimated already saturated. 67 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | text-davinci-003 | 75.91 | #1146 |
| 2 | LLaMA-30B | 75.23 | #1423 |
| 3 | GPT-3.5 Turbo (0613) | 74.77 | #1118 |
| 4 | text-davinci-002 | 73.26 | #1225 |
| 5 | Llama 2 70B | 72.73 | #1201 |
| 6 | mpt-30B | 72.27 | #1476 |
| 7 | Mistral-7B-v0.1 | 70.68 | #1326 |
| 8 | Llama 2 13B | 70.68 | #1436 |
| 9 | LLaMA-65B | 70.23 | #1303 |
| 10 | Palmyra X (43B) | 70.08 | #1317 |
| 11 | RedPajama-INCITE-7B-Instruct | 69.55 | #1566 |
| 12 | vicuna-7B-v1.3 | 69.32 | #1507 |
| 13 | mpt-30B Instruct | 67.95 | #1470 |
| 14 | RedPajama-INCITE-Instruct-3B-v1 | 66.14 | #1617 |
| 15 | falcon-40B | 66.14 | #1442 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=helm-classic-raft · How It Works · Data refreshed daily, snapshot 2026-10-20.