HellaSwag: leaderboard
Commonsense natural language inference benchmark with 70K questions. Given a scenario, models must pick the most plausible continuation. Heavily saturated by modern models.
Metric: Accuracy (%). Source: epoch.ai. Status: saturated. 76 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Median Human | 95.6 |
| 2 | GPT-4 (0314) | 95.3 |
| 3 | Llama 3.1 405B | 89.2 |
| 4 | DeepSeek V3 | 88.9 |
| 5 | DeepSeek V2 | 87.1 |
| 6 | Mixtral 8x7B (v0.1) | 86.7 |
| 7 | text-davinci-003 | 85.5 |
| 8 | Llama 2 70B Base | 85.3 |
| 9 | falcon-40B | 85.28 |
| 10 | Qwen 2.5 72B | 84.8 |
| 11 | LLaMA-65B | 84.2 |
| 12 | Qwen 2.5 Coder 32B | 83 |
| 13 | falcon-11B | 82.91 |
| 14 | Phi-3-medium-128k-instruct | 82.4 |
| 15 | gemma-7B | 82.2 |
Interactive version: theaggregate.ai/benchmark?slug=hellaswag · How It Works · Data refreshed daily, snapshot 2026-09-05.