Sean

### # - [X] I have searched the existing issues ### Is your feature request related to a problem? Please describe it Requesting support for DeepSeek's API, it's really an...

### # - [X] I have searched the existing issues ### Is your feature request related to a problem? Please describe it Requesting support for DeepSeek's API, it's really an...

I'm using -c config.yaml to pass config. When "attributes" is a list of multiple elements: ```bash uv run dolma -c analyze.yaml stat attributes: - cudo/attributes/c4_v2 - cudo/attributes/pii_regex_with_counts_fast_v2 bins: 20 debug:...

Chat template or special tokens are modified during SFT training

2

The SFT training seems to have some problem handling special tokens: I did SFT on Qwen 3 4B but with DeepSeek Qwen 3 8B's special tokens and chat template: ```json...

KD (distillation) chat template error when using pre-collected logprobs

5

### Please check that this issue hasn't been reported before. - [x] I searched previous [Bug Reports](https://github.com/axolotl-ai-cloud/axolotl/labels/bug) didn't find any similar reports. ### Expected Behavior Seems like the `ChatTemplateStrategyWithKDv2` requires:...

Support NVIDIA cuOpt solver

3

## Summary Please consider support NVIDIA cuOpt as a backend solver. It's recently open sourced, with GPU acceleration. NVIDIA has been working with COINOR, so PuLP already can work with...