verl
verl copied to clipboard
[Question] How to use customized chat template
What is the currently recommended method for customizing chat templates? I see there are two approaches, but both have some issues:
- In
rl_dataset.py, it defaults to using the tokenizer's chat template to generate prompts, so one method is to modify the tokenizer's config. - I notice that the
RLHFDatasetinitialization function supportschat_template_func, but this function is not being used and there's no documentation or comments explaining how to define it?