Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, Thomas Scialom
In this work, we develop and release Llama 2, a collection of pretrained and fine-tuned large language models (LLMs) ranging in scale from 7 billion to 70 billion parameters. Our fine-tuned LLMs, called Llama 2-Chat, are optimized for dialogue use cases. Our models outperform open-source chat models on most benchmarks we tested, and based on our human evaluations for helpfulness and safety, may be a suitable substitute for closed-source models. We provide a detailed description of our approach to fine-tuning and safety improvements of Llama 2-Chat in order to enable the community to build on our work and contribute to the responsible development of LLMs.
| Task | Dataset | Metric | Value | Model |
|---|---|---|---|---|
| Transfer Learning | MML | Average (%) | 62.6 | LLaMA 2 34B (5-shot) |
| Transfer Learning | MML | Average (%) | 54.8 | LLaMA 2 13B (5-shot) |
| Transfer Learning | MML | Average (%) | 45.3 | LLaMA 2 7B (5-shot) |
| Question Answering | MultiTQ | Hits@1 | 18.5 | LLaMA2 |
| Question Answering | Natural Questions | EM | 33 | LLaMA 2 70B (one-shot) |
| Question Answering | PIQA | Accuracy | 82.8 | LLaMA 2 70B (0-shot) |
| Question Answering | PIQA | Accuracy | 81.9 | LLaMA 2 34B (0-shot) |
| Question Answering | PIQA | Accuracy | 80.5 | LLaMA 2 13B (0-shot) |
| Question Answering | PIQA | Accuracy | 78.8 | LLaMA 2 7B (0-shot) |
| Question Answering | UniProtQA | BLEU-2 | 0.019 | Llama2-7B-chat |
| Question Answering | UniProtQA | BLEU-4 | 0.002 | Llama2-7B-chat |
| Question Answering | UniProtQA | MEATOR | 0.052 | Llama2-7B-chat |
| Question Answering | UniProtQA | ROUGE-1 | 0.103 | Llama2-7B-chat |
| Question Answering | UniProtQA | ROUGE-2 | 0.06 | Llama2-7B-chat |
| Question Answering | UniProtQA | ROUGE-L | 0.009 | Llama2-7B-chat |
| Question Answering | BoolQ | Accuracy | 85 | LLaMA 2 70B (0-shot) |
| Question Answering | BoolQ | Accuracy | 83.7 | LLaMA 2 34B (0-shot) |
| Question Answering | BoolQ | Accuracy | 81.7 | LLaMA 2 13B (0-shot) |
| Question Answering | BoolQ | Accuracy | 77.4 | LLaMA 2 7B (zero-shot) |
| Question Answering | TriviaQA | EM | 85 | LLaMA 2 70B (one-shot) |
| Question Answering | PubChemQA | BLEU-2 | 0.075 | Llama2-7B-chat |
| Question Answering | PubChemQA | BLEU-4 | 0.009 | Llama2-7B-chat |
| Question Answering | PubChemQA | MEATOR | 0.149 | Llama2-7B-chat |
| Question Answering | PubChemQA | ROUGE-1 | 0.184 | Llama2-7B-chat |
| Question Answering | PubChemQA | ROUGE-2 | 0.043 | Llama2-7B-chat |
| Question Answering | PubChemQA | ROUGE-L | 0.142 | Llama2-7B-chat |
| Question Answering | MMLU (Professional medicine) | Accuracy | 43.38 | Llama2-7B |
| Question Answering | MMLU (Professional medicine) | Accuracy | 40.07 | Llama2-7B-chat |
| Question Answering | MAWPS | Accuracy (%) | 82.4 | LLaMA 2-Chat |
| Question Answering | SVAMP | Execution Accuracy | 69.2 | LLaMA 2-Chat |
| Code Generation | MBPP | Accuracy | 45 | Llama 2 70B (zero-shot) |
| Code Generation | MBPP | Accuracy | 33 | Llama 2 34B (0-shot) |
| Code Generation | MBPP | Accuracy | 30.6 | Llama 2 13B (0-shot) |
| Code Generation | MBPP | Accuracy | 20.8 | Llama 2 7B (0-shot) |
| Math Word Problem Solving | MAWPS | Accuracy (%) | 82.4 | LLaMA 2-Chat |
| Math Word Problem Solving | SVAMP | Execution Accuracy | 69.2 | LLaMA 2-Chat |
| Mathematical Question Answering | MAWPS | Accuracy (%) | 82.4 | LLaMA 2-Chat |
| Mathematical Question Answering | SVAMP | Execution Accuracy | 69.2 | LLaMA 2-Chat |
| Multi-Task Learning | MML | Average (%) | 62.6 | LLaMA 2 34B (5-shot) |
| Multi-Task Learning | MML | Average (%) | 54.8 | LLaMA 2 13B (5-shot) |
| Multi-Task Learning | MML | Average (%) | 45.3 | LLaMA 2 7B (5-shot) |
| Mathematical Reasoning | MAWPS | Accuracy (%) | 82.4 | LLaMA 2-Chat |
| Mathematical Reasoning | SVAMP | Execution Accuracy | 69.2 | LLaMA 2-Chat |
| Sentence Completion | HellaSwag | Accuracy | 85.3 | LLaMA 2 70B (0-shot) |
| Sentence Completion | HellaSwag | Accuracy | 83.3 | LLaMA 2 34B (0-shot) |
| Sentence Completion | HellaSwag | Accuracy | 80.7 | LLaMA 2 13B (0-shot) |
| Sentence Completion | HellaSwag | Accuracy | 77.2 | LLaMA 2 7B (0-shot) |
| Arithmetic Reasoning | GSM8K | Accuracy | 56.8 | LLaMA 2 70B (on-shot) |
| Arithmetic Reasoning | GSM8K | Parameters (Billion) | 70 | LLaMA 2 70B (on-shot) |