HellaSwag (PapersWithCode): leaderboard
HellaSwag (Zellers et al., 2019): pick the most plausible of four endings for an ActivityNet or WikiHow context, with the wrong endings chosen by adversarial filtering. This board is the archived Papers With Code table of accuracies as each paper reported them, under that paper’s own setup, not a common evaluation run; higher is better.
Metric: Accuracy. Source: huggingface.co. Saturation forecast: Rough model projection: around 2026. 90 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | CompassMTL 567M with Tailor | 96.1 | |
| 2 | CompassMTL 567M | 95.6 | |
| 3 | DeBERTa-Large 304M (classification-based) | 95.6 | |
| 4 | GPT-4 (10-shot) | 95.3 | |
| 5 | LLaMA3+MoSLoRA | 95 | |
| 6 | DeBERTa-Large 304M | 94.7 | |
| 7 | LLaMA-2 13B + MixLoRA | 94.7 | |
| 8 | Unicorn 11B (fine-tuned) | 93.9 | |
| 9 | LLaMA-3 8B + MixLoRA | 93.3 | |
| 10 | LLaMA-2 7B + MixLoRA | 93.1 |
Interactive version: theaggregate.ai/benchmark?slug=hellaswag-paperswithcode · How It Works · Data refreshed daily, snapshot 2026-11-05.