Source: LIMA: Less Is More for Alignment

Meta's 2023 paper showing 1,000 curated SFT examples can align a 65B model.

Contents
  • What: research paper (arXiv:2305.11206, NeurIPS 2023)
  • Author: Chunting Zhou, Pengfei Liu, Puxin Xu, et al. (Meta AI and others)
  • Date: 2023-05-18
  • Original: arXiv:2305.11206

Summary

Fine-tunes LLaMA 65B on 1,000 carefully curated prompt-response pairs with plain supervised loss and no RLHF, and finds it competitive with much more heavily aligned systems. Argues that almost all knowledge comes from pre-training.

Key claims

  • 1,000 curated examples, standard SFT loss, no RL or preference modelling.
  • Human raters judged LIMA equivalent to or better than GPT-4 in 43% of cases (58% vs Bard, 65% vs DaVinci003).
  • Learned specific response formats from a handful of examples.
  • Most knowledge is learned in pre-training; alignment mostly teaches format and style.

Feeds

Fine-tuning

Suggest changes →