tocmo0nlord/axolotl
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

feat(doc): update thinking and chat_template notes (#3114) [skip ci]

Browse Source 
* feat: update thinking and chat_template notes

* fix: grammar
This commit is contained in:
NanoCode012
2025-09-12 14:45:18 +07:00
committed by GitHub
parent 9406c0c488
commit fcfc13d710
1 changed files with 10 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
examples/gpt-oss/README.md
View File

@@ -106,6 +106,16 @@ See [Nanobit/text-tools-2k-test](https://huggingface.co/datasets/Nanobit/text-to
Refer to [our docs](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#using-tool-use) for more info. Refer to [our docs](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#using-tool-use) for more info.
### Thinking and chat_template masking conflict
OpenAI’s Harmony template hides `thinking` in all non-final turns, which conflicts with Axolotl’s `chat_template` masking.
If your dataset has `thinking` content mid-turn, there are two paths we recommend:
- Train only on the last turn. This can be accomplished via chat_template's [train on last doc](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#training-on-last-message).
- Adjust your dataset to only have `thinking` content in the last turn.
### TIPS ### TIPS
- Read more on how to load your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html). - Read more on how to load your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html).