AlpacaEval 1.0 — leaderboard
Original instruction-following evaluation using raw win rates against text-davinci-003. 805 prompts scored by an LLM judge. Predecessor to AlpacaEval 2.0.
Metric: Win Rate (%). Source: tatsu-lab.github.io. Status: saturated. 102 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Mistral Medium | 96.83 |
| 2 | GPT-4 | 95.28 |
| 3 | tulu-2-dpo-70B | 95.03 |
| 4 | Mixtral 8x7B Instruct (v0.1) | 94.78 |
| 5 | GPT-4 (0314) | 94.78 |
| 6 | Yi 34B (Chat) | 94.08 |
| 7 | GPT-4 (0613) | 93.78 |
| 8 | Mistral 7B Instruct (v0.2) | 92.78 |
| 9 | Llama 2 70B Chat (HF) | 92.66 |
| 10 | Claude 2 | 91.36 |
| 11 | zephyr-7B-beta | 90.6 |
| 12 | GPT-3.5 Turbo (0301) | 89.37 |
| 13 | WizardLM-13B-V1.2 | 89.17 |
| 14 | vicuna-33B-v1.3 | 88.99 |
| 15 | tulu-2-dpo-13B | 88.12 |
Interactive version: theaggregate.ai/benchmark?slug=alpacaeval-1-0 · How the rankings work · Data refreshed daily, snapshot 2026-07-22.