modelscope / modelscope/ms-swift

Discussion: Does DFT really have practical effects?

Open
#5,386 4 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
15.7k
Forks
1.7k
Avg merge
1d 16h
Merged PRs (30d)
136

Description

I noticed that ms-swift implements DFT by adding --enable_dft_loss true. However, based on my practice, with all other hyperparameters remaining unchanged, merely enabling DFT results in lower loss, but the performance of the model trained with DFT is significantly inferior to that of SFT.

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start by reproducing the SFT-versus-DFT comparison with --enable_dft_loss true, keeping the other hyperparameters unchanged as described. Then trace where this option is handled and document whether the lower loss and inferior model performance are expected; done means the behavior is explained or a concrete correction is identified.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
machine-learning
Issue type
Bug
Difficulty
5/5
Estimated time
Over a week
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.