Inquiry about the SFT

[deadlykitten4]

Hi, thanks for your valuable work!

I notice that the SFT script you provided is full parameter training. Have you tried the lora fine-tuning?

[chengshuang18]

Thank you for your interest and support!

So far, we haven’t tried LoRA fine-tuning yet—the current SFT script is based on full parameter training. We might explore LoRA in the future when scheduling allows, but there’s no concrete plan at the moment.

We warmly welcome anyone from the community to develop LoRA fine-tuning based on this project. If you make progress or gain insights, we’d be glad if you could share them with everyone!

[deadlykitten4]

Hi, @chengshuang18, does the fine-tuning script in this repo also support the SDAR-30BA3B model?

[chengshuang18]

Currently, the repo does not provide a fine-tuning script specifically for the SDAR-30BA3B model. In theory, you can train SDAR-30BA3B by making some modifications to the modeling files based on the existing script for qwen3-30BA3B.

I will be adding a version with support for SDAR-30BA3B modeling files soon and will update the repo accordingly.

[deadlykitten4]

ok! Thank you!