Allow passing args to `apply_chat_template`

[BitPhinix]

⚠️ Please check that this feature request hasn't been suggested before.

• [x] I searched previous Ideas in Discussions didn't find any similar feature requests.
• [x] I searched previous Issues didn't find any similar feature requests.

🔖 Feature description

Allow passing args like enable_thinking to tokizer.apply_chat_template to enable SFT without thinking traces for models like qwen3

✔️ Solution

Either add an option to explicitly set enable_thinking in the config, or add a more generalized way to pass down args to tokizer.apply_chat_template

❓ Alternatives

No response

📝 Additional Context

No response

Acknowledgements

• [x] My issue title is concise, descriptive, and in title casing.
• [x] I have searched the existing issues to make sure this feature has not been requested yet.
• [x] I have provided enough information for the maintainers to understand and evaluate this request.

[NanoCode012]

Hey, this could be an interesting option as model creators expand on that function.

Could you help refresh how enable_thinking help for training? Shouldn't the dataset already contain the thinking traces?

[casper-hansen]

enable_thinking is by default True when using apply_chat_template. That means axolotl is basically incompatible with training Qwen3 as a non-thinking model, which may be desirable for a lot of use-cases where you don't have thinking data and can't get it. Do you think you could help introduce this parameter? @NanoCode012

[NanoCode012]

@casper-hansen , thanks for that refresher. In this case, would a better default be False? I can look into allowing chat_template kwargs

[casper-hansen]

@NanoCode012 I'm not sure of the internals in axolot, but a good check is to figure out where/if apply_chat_template is used and then allow chat_template kwargs.

[NanoCode012]

I linked a draft PR for this. If either of you are interested in testing this, please do give it a try.