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 “2.2 Manually Authored Examples” · whole deck

2.2 Manually Authored Examples

  1. Why did LIMA’s authors write answers in a uniform format?
    Responses typically acknowledge the question, then answer it. Preliminary experiments showed this consistent format helped, perhaps by helping the model form a chain of thought, like “let’s think step by step”.
    • Zhou et al., 2023, LIMA (Advances in Neural Information Processing Systems 36 (NeurIPS 2023)), §2.2
  2. How did LIMA learn to handle harmful requests?
    From only 13 training prompts with some toxicity or malice, answered by partially or fully refusing and explaining why. On 30 sensitive test prompts LIMA responded safely 80% of the time, but implicit malicious intent often slipped through.
    • Zhou et al., 2023, LIMA (Advances in Neural Information Processing Systems 36 (NeurIPS 2023)), §2.2

Source