Add Direct Preference Optimization (DPO) method #1279
+211
−3
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Fixes ml-explore#513 Implement the Direct Preference Optimization (DPO) method as a Reinforcement Learning from Human Feedback (RLHF) example. * **Add DPO Functions**: Add `get_batched_logps` and `dpo_loss` functions to `llms/mlx_lm/utils.py` for DPO implementation. * **Update Training Logic**: Update `llms/mlx_lm/tuner/trainer.py` to include DPO-specific training logic, including a new `dpo_loss` function and condition to check for DPO loss in the training loop. * **Add Configuration Options**: Add configuration options for DPO in `llms/mlx_lm/examples/lora_config.yaml`. * **Update Documentation**: Update `llms/mlx_lm/README.md` to include instructions for using DPO. * **Add Unit Tests**: Add `llms/tests/test_dpo.py` with unit tests for `get_batched_logps`, `dpo_loss`, and DPO-specific training logic. --- For more details, open the [Copilot Workspace session](https://copilot-workspace.githubnext.com/ml-explore/mlx-examples/issues/513?shareId=XXXX-XXXX-XXXX-XXXX).
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Fixes #513
Implement the Direct Preference Optimization (DPO) method as a Reinforcement Learning from Human Feedback (RLHF) example.
get_batched_logpsanddpo_lossfunctions tollms/mlx_lm/utils.pyfor DPO implementation.llms/mlx_lm/tuner/trainer.pyto include DPO-specific training logic, including a newdpo_lossfunction and condition to check for DPO loss in the training loop.llms/mlx_lm/examples/lora_config.yaml.llms/mlx_lm/README.mdto include instructions for using DPO.llms/tests/test_dpo.pywith unit tests forget_batched_logps,dpo_loss, and DPO-specific training logic.For more details, open the Copilot Workspace session.