Spaces:

---

Fourwheels2512

/

crma-fine-tuner

like 0

Sleeping

Thanks for visiting CRMA Fine-Tuner!

This Space lets you fine-tune TinyLlama, Gemma 2B, and Mistral-7B on your own data with built-in training stability (CRMA + ZClip).

How to use

1. Upload a CSV or JSONL file with instruction and output columns
2. Choose your model (TinyLlama for fast tests, Gemma or Mistral for better quality)
3. Hit Run and monitor live loss + gradient norm charts
4. Download your adapter ZIP when complete — includes merged weights + CRMA adapter + injector script

What makes this different from plain LoRA?

CRMA (Constrained Residual Mixing Adapter) adds a Sinkhorn-constrained mixing layer alongside LoRA. In our ablations:

• TinyLlama: peak gradient norm -52.7% vs plain LoRA, same final loss
• Mistral-7B: plain LoRA hit a catastrophic spike (grad norm ~263); CRMA held it at ~3 (-98.9%)

Questions or feedback?

Drop a comment here — happy to help with dataset formatting, model selection, or anything else.