learn.ifs.name

LIMA: Less Is More for Alignment

Zhou et al., 2023, LIMA (Advances in Neural Information Processing Systems 36 (NeurIPS 2023))

Subscriber deck — free during early access

2 cards in “4.2 Results” · whole deck

4.2 Results

  1. How did LIMA compare with other assistants in pairwise human preferences?
    Over 300 test prompts it was preferred to Alpaca 65B (trained on 52× more data) and to RLHF-trained DaVinci003; against Bard, GPT-4 and Claude it was equal or better in 58%, 43% and 46% of cases. GPT-4 as judge showed the same trends, even preferring LIMA to its own output 19% of the time.
    • Zhou et al., 2023, LIMA (Advances in Neural Information Processing Systems 36 (NeurIPS 2023)), §4.2
  2. How did GPT-4 compare with human annotators when judging LIMA’s outputs?
    Using tie-discounted accuracy, GPT-4 agreed with crowd workers 78% and with authors 79% of the time, about as often as humans agreed with each other (78–82%).
    • Zhou et al., 2023, LIMA (Advances in Neural Information Processing Systems 36 (NeurIPS 2023)), §4.2

Source