search

locuslab / tofu

Landing Page for TOFU
MIT License
68 stars 15 forks source link

Why finetuned model and retained model have similar model utility? #37

Closed Carol-gutianle closed 3 days ago

Carol-gutianle commented 3 days ago

When I reproduced ToFU's performance on llama2-finetuned(origin) and llama2-chat(retain), I found that there was a significant difference in model utility between the two. However, in the leaderboard, these values are almost identical. If an unfinetuned model can also perform so well on ToFU, does this imply some degree of data leakage? I look forward to your response. Thank you for your time.

  1. my result：llama-chat(finetune) on each forget split all gets the Model Utility of 0.2487, and retained version gets the Model Uitlity of 0.624.
  2. results on leaderboard: image
molereddy commented 3 days ago

Retain model is not unfinetuned. It is trained with retain split. Finetune model is trained with full TOFU split.

Carol-gutianle commented 3 days ago

Thank you for your explanation.