Knowledge
HellaSwag Leaderboard
HellaSwag is a commonsense benchmark that asks a model to pick the most plausible continuation of an everyday situation. The wrong options are adversarially chosen to fool models, testing grounded commonsense understanding.
Source: epoch42 open models ranked+34 proprietaryData through Dec 2024
All models ranked on HellaSwag
Proprietary / closed models are shown dimmed — you can't run them locally, but they show where the open field stands.
| # | Model | Score |
|---|---|---|
| 1 | GPT 4 (Mar 14) · proprietary | 95.3% |
| 2 | GPT 4 32K (Mar 14) · proprietary | 95.3% |
| 3 | Llama 3.1 405B · 405.9B | 89.2% |
| 4 | Falcon 180B · 180B | 89.0% |
| 5 | DeepSeek v3 · 684.5B | 88.9% |
| 6 | DeepSeek v2 · 235.7B | 87.1% |
| 7 | PaLM 2 L · proprietary | 86.8% |
| 8 | Mixtral 8x7B v0.1 · 46.7B | 86.7% |
| 9 | Text Davinci 003 · proprietary | 85.5% |
| 10 | Llama 2 70B HF · 69.0B | 85.3% |
| 11 | Falcon 40B · 41.8B | 85.3% |
| 12 | Qwen2.5 72B · 72.7B | 84.8% |
| 13 | Llama 65B · proprietary | 84.2% |
| 14 | StableBeluga2 · 70B | 84.1% |
| 15 | PaLM 2 M · proprietary | 84.0% |
| 16 | PaLM 540B · proprietary | 83.8% |
| 17 | Qwen2.5 Coder 32B · 32.8B | 83.0% |
| 18 | Falcon 11B · 11.1B | 82.9% |
| 19 | Llama 33B · proprietary | 82.8% |
| 20 | Megatron Turing NLG 530B · proprietary | 82.4% |
| 21 | Nemotron 4 15B · proprietary | 82.4% |
| 22 | Phi 3 Medium 128K Instruct · proprietary | 82.4% |
| 23 | Gemma 7B · 8.5B | 82.2% |
| 24 | PaLM 2 S · proprietary | 82.0% |
| 25 | Text Davinci 002 · proprietary | 81.5% |
| 26 | Mistral 7B v0.1 · 7B | 81.0% |
| 27 | Llama 2 13B HF · 13.0B | 80.7% |
| 28 | Qwen2.5 Coder 14B · 14.8B | 80.2% |
| 29 | Text Davinci 001 · proprietary | 79.3% |
| 30 | Gopher (280B) · proprietary | 79.2% |
| 31 | Llama 13B · proprietary | 79.2% |
| 32 | Opt 175B · proprietary | 79.1% |
| 33 | Falcon 7B · 7.2B | 78.1% |
| 34 | Internlm 20B · 20B | 78.1% |
| 35 | Davinci · proprietary | 77.5% |
| 36 | GLaM (MoE) · proprietary | 77.2% |
| 37 | Llama 2 7B HF · 6.7B | 77.2% |
| 38 | Phi 3 Small 8k Instruct · 7.4B | 77.0% |
| 39 | Qwen2.5 Coder 7B · 7.6B | 76.8% |
| 40 | Phi 3 Mini 4k Instruct · 3.8B | 76.7% |
| 41 | Mpt 7B · proprietary | 76.4% |
| 42 | Yi 9B · 8.8B | 76.4% |
| 43 | Llama 7B · 6.7B | 76.2% |
| 44 | Opt 66B · proprietary | 74.5% |
| 45 | Bloom · 176.2B | 74.4% |
| 46 | Yi 6B · 6.1B | 74.4% |
| 47 | Xgen 7B 8k Base · 7B | 74.2% |
| 48 | Open Llama 7B · proprietary | 71.8% |
| 49 | INTELLECT 1 Instruct · 10.2B | 71.4% |
| 50 | Gemma 2B · 2.5B | 71.4% |
| 51 | Qwen2.5 Coder 3B · 3.1B | 70.9% |
| 52 | Baichuan2 13B Base · 13B | 70.8% |
| 53 | Dolly v2 12B · proprietary | 70.8% |
| 54 | Internlm 7B · 7B | 70.6% |
| 55 | GPT Neox 20B · 20.7B | 70.5% |
| 56 | RedPajama INCITE 7B Base · proprietary | 70.3% |
| 57 | Opt 13B · proprietary | 69.9% |
| 58 | Curie · proprietary | 68.2% |
| 59 | Baichuan2 7B Base · 7B | 68.0% |
| 60 | Text Curie 001 · proprietary | 67.6% |
| 61 | GPT J 6B · 6B | 66.2% |
| 62 | Qwen2.5 Coder 1.5B · 1.5B | 61.8% |
| 63 | Cerebras GPT 13B · 13B | 59.4% |
| 64 | Vicuna 13B v1.1 · proprietary | 57.8% |
| 65 | Chatglm2 6B · 6B | 57.0% |
| 66 | Text Babbage 001 · proprietary | 56.1% |
| 67 | Babbage · proprietary | 55.5% |
| 68 | Phi 2 · 2.8B | 53.6% |
| 69 | Qwen2.5 Coder 0.5B · 494M | 48.4% |
| 70 | Phi 1 5 · 1.4B | 47.6% |
| 71 | Ada · proprietary | 43.5% |
| 72 | Text Ada 001 · proprietary | 42.9% |
| 73 | GPT Neo 2.7B · proprietary | 42.7% |
| 74 | Opt 1.3b · proprietary | 41.5% |
| 75 | Stablelm Tuned Alpha 7B · 7B | 40.7% |
| 76 | Gpt2 Xl · 1.6B | 40.0% |
Score vs model size
Which models give the most quality for their size — the ones worth running locally.
- Qwen2.5 Coder 0.5B, 494M, score 48.4% — on the efficiency frontier (best score at its size or smaller).
- Qwen2.5 Coder 1.5B, 2B, score 61.8% — on the efficiency frontier (best score at its size or smaller).
- Gemma 2B, 3B, score 71.4% — on the efficiency frontier (best score at its size or smaller).
- Phi 3 Mini 4k Instruct, 4B, score 76.7% — on the efficiency frontier (best score at its size or smaller).
- Llama 2 7B HF, 7B, score 77.2% — on the efficiency frontier (best score at its size or smaller).
- Mistral 7B v0.1, 7B, score 81.0% — on the efficiency frontier (best score at its size or smaller).
- Gemma 7B, 9B, score 82.2% — on the efficiency frontier (best score at its size or smaller).
- Falcon 11B, 11B, score 82.9% — on the efficiency frontier (best score at its size or smaller).
- Qwen2.5 Coder 32B, 33B, score 83.0% — on the efficiency frontier (best score at its size or smaller).
- Falcon 40B, 42B, score 85.3% — on the efficiency frontier (best score at its size or smaller).
- Mixtral 8x7B v0.1, 47B, score 86.7% — on the efficiency frontier (best score at its size or smaller).
- Falcon 180B, 180B, score 89.0% — on the efficiency frontier (best score at its size or smaller).
- Llama 3.1 405B, 406B, score 89.2% — on the efficiency frontier (best score at its size or smaller).
HellaSwag: frequently asked questions
- What is the best open LLM on HellaSwag?
- Llama 3.1 405B is the top open model on HellaSwag, scoring 89.2%. Among all models tested — including proprietary ones — it ranks #3. The top model overall is GPT 4 (Mar 14) (OpenAI) at 95.3%.
- What's the best HellaSwag model you can run on a 24 GB GPU?
- Falcon 40B is the highest-scoring open model that fits in 24 GB at 4-bit quantization (about 23 GB), scoring 85.3% on HellaSwag.
- What's the best HellaSwag model you can run on a 12 GB GPU?
- Falcon 11B is the highest-scoring open model that fits in 12 GB at 4-bit quantization (about 6 GB), scoring 82.9% on HellaSwag.
- Can open models match proprietary models on HellaSwag?
- Not quite on HellaSwag: the strongest proprietary model (GPT 4 (Mar 14)) scores 95.3%, ahead of the best open model (Llama 3.1 405B) at 89.2% — but you can run the open one yourself.
Scores aggregated from epoch. llmrun does not run this benchmark — see the source for methodology, or the about benchmarks for what it measures.