Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Contrastive Error Attribution for Finetuned Language Models #2116

Open
icoxfog417 opened this issue Nov 7, 2023 · 0 comments
Open

Contrastive Error Attribution for Finetuned Language Models #2116

icoxfog417 opened this issue Nov 7, 2023 · 0 comments

Comments

@icoxfog417
Copy link
Member

一言でいうと

文書生成においてハルシネーションを引き起こすデータセット内のデータを高精度で特定する手法の提案。既存の手法はエラーケースとの類似度を推論時の勾配から推定していたが、本来は正常な生成をエラーより重みづけしたいため正常とエラートの勾配の差分をもとに判断する手法を提案。要約データセットでBARTを学習し、エラー除去後データで Fine Tune することで 70% の抑制に成功 (既存手法 20% 台)。

論文リンク

https://arxiv.org/abs/2212.10722v2

著者/所属機関

Faisal Ladhak, Esin Durmus, Tatsunori Hashimoto

  • Columbia University
  • Stanford University

投稿日付(yyyy/MM/dd)

2022/12/21

概要

新規性・差分

手法

結果

コメント

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

No branches or pull requests

1 participant