---
base_model:
- qingy2024/Qwarkstar-4B
license: apache-2.0
datasets:
- HuggingFaceTB/smoltalk
language:
- en
pipeline_tag: text-generation
---

## Qwarkstar 4B Instruct (Preview)

> [!NOTE]
> Training complete!

This model is fine-tuned using Supervised Fine-Tuning (SFT) on 100k samples from the `HuggingFaceTB/smoltalk` dataset.  
It follows the ChatML input-output formatting template.  

### Training Details:
- **Base Model**: `qingy2024/Qwarkstar-4B`
- **Batch Size**: 32 (2 H100s x 8 per GPU)
- **Max Gradient Norm**: 1.0
- **Final Loss**: ~0.59