HellaSwag
ReportedChoosing the most sensible way to continue a short description of an everyday situation.
- Kind
- Published by a third party's leaderboard, collected by Epoch AI. The leaderboard is named and linked.
- Score shown
- Epoch's column "Overall accuracy", as a percentage, higher is better; random guessing scores about 25.0%. The source gives no error bars.
- Results
- 113 in all: 71 for open-weight models, 28 for closed ones, 14 where Epoch does not say. 29 linked to a model page here.
Best open-weight result: Llama-3.1-405B, 89.2%, number 3 of 113 results.
| # | Model | Weights | Setting | Score | Date | Source |
|---|---|---|---|---|---|---|
| 1 | gpt-4-0314OpenAI | closed | Shots: 10 | 95.3% | not dated | GPT-4 Technical Report |
| 2 | gpt-4-32k-0314OpenAI | closed | Shots: 10 | 95.3% | not dated | The Falcon Series of Open Language Models |
| 3 | Llama-3.1-405BMeta AIno page match | open | Shots: 10 | 89.2% | not dated | DeepSeek-V3 Technical Report |
| 4 | falcon-180BTechnology Innovation Instituteno page match | open | Shots: 10 | 89.0% | not dated | The Falcon Series of Open Language Models |
| 5 | deepseek-ai/DeepSeek-V3Epoch: DeepSeek-V3, DeepSeek | open | Shots: 10 | 88.9% | not dated | DeepSeek-V3 Technical Report |
| 6 | DeepSeek-V2DeepSeekno page match | open | Shots: 10 | 87.1% | not dated | DeepSeek-V3 Technical Report |
| 7 | PaLM 2-L | not stated | Shots: 1 | 86.8% | not dated | PaLM 2 Technical Report |
| 8 | mistralai/Mixtral-8x7B-v0.1Epoch: Mixtral-8x7B-v0.1, Mistral AI | open | Shots: 10 | 86.7% | not dated | Mixtral of Experts |
| 9 | text-davinci-003OpenAI | not stated | Shots: 10 | 85.5% | not dated | The Falcon Series of Open Language Models |
| 10 | text-davinci-003OpenAI | not stated | Shots: 10 | 85.5% | not dated | GPT-4 Technical Report |
| 11 | meta-llama/Llama-2-70b-hfEpoch: Llama-2-70b-hf, Meta AI | open | Shots: 0 | 85.3% | not dated | Qwen Technical Report |
| 12 | falcon-40bTechnology Innovation Instituteno page match | open | Shots: 10 | 85.3% | not dated | Falcon2-11B Technical Report |
| 13 | falcon-40bTechnology Innovation Instituteno page match | open | Shots: 10 | 85.3% | not dated | Falcon2-11B Technical Report |
| 14 | Qwen2.5-72BAlibabano page match | open | Shots: 10 | 84.8% | not dated | DeepSeek-V3 Technical Report |
| 15 | mistralai/Mixtral-8x7B-v0.1Epoch: Mixtral-8x7B-v0.1, Mistral AI | open | default | 84.4% | not dated | Mixtral of Experts |
| 16 | LLaMA-65BMeta AIno page match | open | Shots: 0 | 84.2% | not dated | LLaMA: Open and Efficient Foundation Language Models |
| 17 | stabilityai/StableBeluga2Epoch: StableBeluga2, Stability AI | open | Shots: 0 | 84.1% | not dated | Qwen Technical Report |
| 18 | PaLM 2-M | not stated | Shots: 1 | 84.0% | not dated | PaLM 2 Technical Report |
| 19 | PaLM 540BGoogle Research | closed | Shots: few | 83.8% | not dated | PaLM: Scaling Language Modeling with Pathways |
| 20 | PaLM 540BGoogle Research | closed | Shots: 1 | 83.6% | not dated | PaLM: Scaling Language Modeling with Pathways |
| 21 | PaLM 540BGoogle Research | closed | Shots: 0 | 83.4% | not dated | PaLM: Scaling Language Modeling with Pathways |
| 22 | Qwen2.5-Coder-32BAlibabano page match | open | default | 83.0% | not dated | Qwen2.5-Coder Technical Report |
| 23 | tiiuae/falcon-11BEpoch: falcon-11b, Technology Innovation Institute | open | Shots: 10 | 82.9% | not dated | Falcon2-11B Technical Report |
| 24 | tiiuae/falcon-11BEpoch: falcon-11b, Technology Innovation Institute | open | Shots: 10 | 82.9% | not dated | Falcon2-11B Technical Report |
| 25 | falcon-40bTechnology Innovation Instituteno page match | open | Shots: 0 | 82.8% | not dated | Falcon2-11B Technical Report |
| 26 | falcon-40bTechnology Innovation Instituteno page match | open | Shots: 0 | 82.8% | not dated | Falcon2-11B Technical Report |
| 27 | LLaMA-33BMeta AIno page match | open | Shots: 0 | 82.8% | not dated | LLaMA: Open and Efficient Foundation Language Models |
| 28 | Megatron-Turing NLG 530BMicrosoft,Nvidia | closed | default | 82.4% | not dated | GLaM: Efficient Scaling of Language Models with Mixture-of-Experts |
| 29 | Megatron-Turing NLG 530BMicrosoft,Nvidia | closed | Shots: few | 82.4% | not dated | Using DeepSpeed and Megatron to Train Megatron-Turing NLG 530B, A Large-Scale Generative Language Model |
| 30 | Nemotron-4 15BNvidia | closed | Shots: 0 | 82.4% | not dated | Nemotron-4 15B Technical Report |
| 31 | microsoft/Phi-3-medium-128k-instructEpoch: Phi-3-medium-128k-instruct, Microsoft | open | Shots: 5 | 82.4% | not dated | Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone |
| 32 | gemma-7bGoogle DeepMindno page match | open | default | 82.2% | not dated | Gemma: Open Models Based on Gemini Research and Technology |
| 33 | text-davinci-003OpenAI | not stated | Shots: 0 | 82.2% | not dated | Stanford HELM |
| 34 | tiiuae/falcon-11BEpoch: falcon-11b, Technology Innovation Institute | open | Shots: 0 | 82.1% | not dated | Falcon2-11B Technical Report |
| 35 | tiiuae/falcon-11BEpoch: falcon-11b, Technology Innovation Institute | open | Shots: 0 | 82.1% | not dated | Falcon2-11B Technical Report |
| 36 | PaLM 2-S | not stated | Shots: 1 | 82.0% | not dated | PaLM 2 Technical Report |
| 37 | text-davinci-002OpenAI | closed | Shots: 0 | 81.5% | not dated | Stanford HELM |
| 38 | gemma-7bGoogle DeepMindno page match | open | Shots: 0 | 81.2% | not dated | Gemma: Open Models Based on Gemini Research and Technology |
| 39 | mistralai/Mistral-7B-v0.1Epoch: Mistral-7B-v0.1, Mistral AI | open | default | 81.0% | not dated | Mixtral of Experts |
| 40 | Llama-2-13bMeta AIno page match | open | Shots: 0 | 80.7% | not dated | Gemma: Open Models Based on Gemini Research and Technology |
| 41 | Megatron-Turing NLG 530BMicrosoft,Nvidia | closed | Shots: 0 | 80.2% | not dated | Using DeepSpeed and Megatron to Train Megatron-Turing NLG 530B, A Large-Scale Generative Language Model |
| 42 | Megatron-Turing NLG 530BMicrosoft,Nvidia | closed | Shots: 1 | 80.2% | not dated | Using DeepSpeed and Megatron to Train Megatron-Turing NLG 530B, A Large-Scale Generative Language Model |
| 43 | Qwen/Qwen2.5-Coder-14BEpoch: Qwen2.5-Coder-14B | open | default | 80.2% | not dated | Qwen2.5-Coder Technical Report |
| 44 | text-davinci-001OpenAI | closed | Shots: few | 79.3% | not dated | Using DeepSpeed and Megatron to Train Megatron-Turing NLG 530B, A Large-Scale Generative Language Model |
| 45 | text-davinci-001OpenAI | closed | Shots: 20 | 79.3% | not dated | GLaM: Efficient Scaling of Language Models with Mixture-of-Experts |
| 46 | text-davinci-001OpenAI | closed | Shots: few | 79.3% | not dated | Language Models are Few-Shot Learners |
| 47 | Gopher (280B)DeepMind | closed | Shots: 0 | 79.2% | not dated | Using DeepSpeed and Megatron to Train Megatron-Turing NLG 530B, A Large-Scale Generative Language Model |
| 48 | Gopher (280B)DeepMind | closed | Shots: 0 | 79.2% | not dated | GLaM: Efficient Scaling of Language Models with Mixture-of-Experts |
| 49 | LLaMA-13BMeta AIno page match | open | Shots: 0 | 79.2% | not dated | LLaMA: Open and Efficient Foundation Language Models |
| 50 | opt-175bMeta AIno page match | open | Shots: 0 | 79.1% | not dated | Stanford HELM |
| 51 | text-davinci-001OpenAI | closed | Shots: 0 | 78.9% | not dated | Using DeepSpeed and Megatron to Train Megatron-Turing NLG 530B, A Large-Scale Generative Language Model |
| 52 | text-davinci-001OpenAI | closed | Shots: 0 | 78.9% | not dated | GLaM: Efficient Scaling of Language Models with Mixture-of-Experts |
| 53 | text-davinci-001OpenAI | closed | Shots: 0 | 78.9% | not dated | Language Models are Few-Shot Learners |
| 54 | falcon-7bTechnology Innovation Instituteno page match | open | Shots: 10 | 78.1% | not dated | Falcon2-11B Technical Report |
| 55 | falcon-7bTechnology Innovation Instituteno page match | open | Shots: 10 | 78.1% | not dated | Falcon2-11B Technical Report |
| 56 | internlm/internlm-20bEpoch: internlm-20b | open | Shots: 0 | 78.1% | not dated | Qwen Technical Report |
| 57 | text-davinci-001OpenAI | closed | Shots: 1 | 78.1% | not dated | Using DeepSpeed and Megatron to Train Megatron-Turing NLG 530B, A Large-Scale Generative Language Model |
| 58 | text-davinci-001OpenAI | closed | Shots: 1 | 78.1% | not dated | GLaM: Efficient Scaling of Language Models with Mixture-of-Experts |
| 59 | text-davinci-001OpenAI | closed | Shots: 1 | 78.1% | not dated | Language Models are Few-Shot Learners |
| 60 | davinciOpenAI | closed | Shots: 0 | 77.5% | not dated | Stanford HELM |
| 61 | GLaM (MoE)Google | closed | Shots: 8 | 77.2% | not dated | GLaM: Efficient Scaling of Language Models with Mixture-of-Experts |
| 62 | Llama-2-7bMeta AIno page match | open | Shots: 0 | 77.2% | not dated | Gemma: Open Models Based on Gemini Research and Technology |
| 63 | microsoft/Phi-3-small-8k-instructEpoch: Phi-3-small-8k-instruct, Microsoft | open | Shots: 5 | 77.0% | not dated | Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone |
| 64 | GLaM (MoE)Google | closed | Shots: 1 | 76.8% | not dated | GLaM: Efficient Scaling of Language Models with Mixture-of-Experts |
| 65 | Qwen2.5-Coder-7BAlibabano page match | open | default | 76.8% | not dated | Qwen2.5-Coder Technical Report |
| 66 | microsoft/Phi-3-mini-4k-instructEpoch: Phi-3-mini-4k-instruct, Microsoft | open | Shots: 5 | 76.7% | not dated | Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone |
| 67 | GLaM (MoE)Google | closed | Shots: 0 | 76.6% | not dated | GLaM: Efficient Scaling of Language Models with Mixture-of-Experts |
| 68 | falcon-7bTechnology Innovation Instituteno page match | open | Shots: 0 | 76.4% | not dated | XGen-7B Technical Report |
| 69 | mpt-7bMosaicMLno page match | open | Shots: 0 | 76.4% | not dated | Qwen Technical Report |
| 70 | 01-ai/Yi-9BEpoch: Yi-9B | open | default | 76.4% | not dated | Yi: Open Foundation Models by 01.AI |
| 71 | falcon-7bTechnology Innovation Instituteno page match | open | Shots: 0 | 76.3% | not dated | Falcon2-11B Technical Report |
| 72 | falcon-7bTechnology Innovation Instituteno page match | open | Shots: 0 | 76.3% | not dated | Falcon2-11B Technical Report |
| 73 | LLaMA-7BMeta AIno page match | open | Shots: 0 | 76.2% | not dated | XGen-7B Technical Report |
| 74 | LLaMA-7BMeta AIno page match | open | Shots: 0 | 76.1% | not dated | LLaMA: Open and Efficient Foundation Language Models |
| 75 | mpt-7bMosaicMLno page match | open | Shots: 0 | 76.1% | not dated | XGen-7B Technical Report |
| 76 | facebook/opt-66bEpoch: opt-66b, Meta AI | open | Shots: 0 | 74.5% | not dated | Stanford HELM |
| 77 | bigscience/bloomEpoch: bloom, Hugging Face,BigScience | open | Shots: 0 | 74.4% | not dated | Stanford HELM |
| 78 | Yi-6B01.AIno page match | open | Shots: 10 | 74.4% | not dated | Yi: Open Foundation Models by 01.AI |
| 79 | Salesforce/xgen-7b-8k-baseEpoch: xgen-7b-8k-base, Salesforce | open | Shots: 0 | 74.2% | not dated | XGen-7B Technical Report |
| 80 | falcon-7bTechnology Innovation Instituteno page match | open | Shots: 0 | 74.1% | not dated | Qwen Technical Report |
| 81 | open_llama_7b | not stated | Shots: 0 | 71.8% | not dated | XGen-7B Technical Report |
| 82 | INTELLECT-1-InstructPrime Intellect,Hugging Face,Arcee AI | not stated | Shots: 10 | 71.4% | not dated | INTELLECT-1 Technical Report |
| 83 | gemma-2bGoogle DeepMindno page match | open | Shots: 0 | 71.4% | not dated | Gemma: Open Models Based on Gemini Research and Technology |
| 84 | Qwen/Qwen2.5-Coder-3BEpoch: Qwen2.5-Coder-3B | open | default | 70.9% | not dated | Qwen2.5-Coder Technical Report |
| 85 | Baichuan-2-13B-BaseBaichuanno page match | open | Shots: 0 | 70.8% | not dated | Qwen Technical Report |
| 86 | dolly-v2-12bDatabricksno page match | open | Shots: 0 | 70.8% | not dated | XGen-7B Technical Report |
| 87 | internlm/internlm-7bEpoch: internlm-7b | open | Shots: 0 | 70.6% | not dated | Qwen Technical Report |
| 88 | EleutherAI/gpt-neox-20bEpoch: gpt-neox-20b, EleutherAI | open | Shots: 0 | 70.5% | not dated | XGen-7B Technical Report |
| 89 | RedPajama-INCITE-7B-Base | not stated | Shots: 0 | 70.3% | not dated | XGen-7B Technical Report |
| 90 | facebook/opt-13bEpoch: opt-13b | open | Shots: 0 | 69.9% | not dated | XGen-7B Technical Report |
| 91 | curieOpenAI | not stated | Shots: 0 | 68.2% | not dated | Stanford HELM |
| 92 | Baichuan-2-7B-BaseBaichuanno page match | open | Shots: 0 | 68.0% | not dated | Qwen Technical Report |
| 93 | text-curie-001OpenAI | closed | Shots: 0 | 67.6% | not dated | Stanford HELM |
| 94 | gpt-j-6bEleutherAI,LAIONno page match | open | Shots: 0 | 66.2% | not dated | XGen-7B Technical Report |
| 95 | Qwen2.5-Coder-1.5BAlibabano page match | open | default | 61.8% | not dated | Qwen2.5-Coder Technical Report |
| 96 | Cerebras-GPT-13BCerebras Systemsno page match | open | Shots: 0 | 59.4% | not dated | XGen-7B Technical Report |
| 97 | vicuna-13b-v1.1 | not stated | default | 57.8% | not dated | Textbooks Are All You Need II: phi-1.5 technical report |
| 98 | Llama-2-7bMeta AIno page match | open | default | 57.1% | not dated | Textbooks Are All You Need II: phi-1.5 technical report |
| 99 | mpt-7bMosaicMLno page match | open | default | 57.1% | not dated | Textbooks Are All You Need II: phi-1.5 technical report |
| 100 | zai-org/chatglm2-6bEpoch: chatglm2-6b | open | Shots: 0 | 57.0% | not dated | Qwen Technical Report |
| 101 | LLaMA-7BMeta AIno page match | open | default | 56.2% | not dated | Textbooks Are All You Need II: phi-1.5 technical report |
| 102 | text-babbage-001OpenAI | closed | Shots: 0 | 56.1% | not dated | Stanford HELM |
| 103 | babbageOpenAI | not stated | Shots: 0 | 55.5% | not dated | Stanford HELM |
| 104 | falcon-7bTechnology Innovation Instituteno page match | open | default | 54.2% | not dated | Textbooks Are All You Need II: phi-1.5 technical report |
| 105 | microsoft/phi-2Epoch: phi-2, Microsoft | open | Shots: 5 | 53.6% | not dated | Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone |
| 106 | Qwen/Qwen2.5-Coder-0.5BEpoch: Qwen2.5-Coder-0.5B | open | default | 48.4% | not dated | Qwen2.5-Coder Technical Report |
| 107 | microsoft/phi-1_5Epoch: phi-1_5, Microsoft | open | Shots: 0 | 47.6% | not dated | Textbooks Are All You Need II: phi-1.5 technical report |
| 108 | adaOpenAI | not stated | Shots: 0 | 43.5% | not dated | Stanford HELM |
| 109 | text-ada-001OpenAI | not stated | Shots: 0 | 42.9% | not dated | Stanford HELM |
| 110 | gpt-neo-2.7BEleutherAIno page match | open | default | 42.7% | not dated | Textbooks Are All You Need II: phi-1.5 technical report |
| 111 | facebook/opt-1.3bEpoch: opt-1.3b, Meta AI | open | default | 41.5% | not dated | Textbooks Are All You Need II: phi-1.5 technical report |
| 112 | stabilityai/stablelm-tuned-alpha-7bEpoch: stablelm-tuned-alpha-7b | open | Shots: 0 | 40.7% | not dated | XGen-7B Technical Report |
| 113 | gpt2-xlOpenAIno page match | open | default | 40.0% | not dated | Textbooks Are All You Need II: phi-1.5 technical report |
Open-weight rows are shaded. Dates are when the result was run or evaluated, where the data says. Data from Epoch AI (CC-BY 4.0), fetched 19 Nov 2026: Epoch AI, ‘Capabilities & benchmarking’. Published online at epoch.ai. Retrieved from ‘https://epoch.ai/benchmarks’ [online resource]. Epoch AI. Method