Source: The False Promise of Imitating Proprietary LLMs

Berkeley 2023 paper finding that fine-tuning open models on ChatGPT outputs copies style, not capability or factuality.

Contents
  • What: research paper (arXiv:2305.15717)
  • Author: Arnav Gudibande, Eric Wallace, Charlie Snell, Xinyang Geng, Hao Liu, Pieter Abbeel, Sergey Levine, Dawn Song
  • Date: 2023-05-25
  • Original: arXiv:2305.15717

Summary

The authors fine-tuned open models (1.5B–13B) on ChatGPT outputs. Crowd raters judged the imitators competitive, but targeted benchmarks showed the gap barely closed on tasks the imitation data did not cover.

Key claims

  • Imitation models copy ChatGPT’s style, not its factuality, and this fools human raters.
  • Little of the capability gap closes on tasks outside the imitation data’s support.
  • Closing the gap by imitation would take impractical amounts of data. Stronger base models are the better lever.

Feeds

Synthetic data generation

Suggest changes →