LLaMA: Open and Efficient Foundation Language Models

We introduce LLaMA, a collection of foundation language models ranging from 7B to 65B parameters. In particular, LLaMA-13B outperforms GPT-3 (175B) on most benchmarks, and LLaMA-65B is competitive with the best models, Chinchilla-70B and PaLM-540B.

⚡ This is an original paraphrased summary — not copied from the abstract. Full paper available at the source link below.

• 1 We train our models on trillions of tokens, and show that it is possible to train advanced models using publicly available datasets exclusively, without resorting to proprietary and inaccessible datasets.
• 2 In particular, LLaMA-13B outperforms GPT-3 (175B) on most benchmarks, and LLaMA-65B is competitive with the best models, Chinchilla-70B and PaLM-540B.
• 3 We release all our models to the research community.

This research advances how AI systems learn, reason, and solve problems — with direct implications for automation and scientific discovery.

This summary is based on publicly available metadata and abstract. For the full research paper, visit the original source: