Add a chat_template prompt strategy for DPO
#1725
Conversation
This mimics the sft chat_template strategy such that users can: * Specify the messages field * Specify the per message role and content fields * speicfy the chosen and rejected fields * Let the tokenizer construct the raw prompt * Ensure the chosen and rejected fields don't have any prefix tokens
| @@ -62,7 +62,7 @@ def process_tokens_for_rl_debug(tokens, color, tokenizer, text_only): | |||
| """Helper function to process and color tokens.""" | |||
| colored_tokens = [ | |||
| color_token_for_rl_debug(tokenizer.decode(token), token, color, text_only) | |||
| for token in tokenizer.encode(tokens) | |||
| for token in tokenizer.encode(tokens, add_special_tokens=False) | |||
There was a problem hiding this comment.
Note: I added this since by default I saw that this step was including the bos token all the time. Since that's already included it seemed reasonable to not add it in a second time.
|
@fozziethebeat but for DPO training, since trl handles the tokenization, do we need this piece? |
Was this in reference to the change in the debugging output? If so, it's not required but I think anyone manually inspecting tokenization output (like i did) would be very surprised to see the bos token duplicated in numerous scenarios. So it's more to give confidence that we constructed the strings correctly. |
|
Any other changes to add before updating the branch and approving for merging? |
Description
Replicates the
chat_templatesupport from SFT datasets but for DPO training. Users can now specify a dataset with a list of conversation messages along with rejected and chosen columns having a single conversation message. Further, all fields can be customized.Motivation and Context
This change provides a more configurable set of datasets for DPO training.
Fixes #1708
How has this been tested?
Screenshots (if appropriate)
Types of changes
Social Handles (Optional)
@fozziethebeat