LLM Fine-tuning (LoRA / QLoRA)
GPU required for training. LoRA parameter count and quantization config cells run on CPU.
Pitfalls
- LoRA rank vs quality: Higher rank (r=64) trains more parameters but is not always better — r=16 is the strong default. Benefit of higher rank depends on dataset size and task diversity.
- Quantization ≠ adapter precision: NF4 quantizes the frozen base model weights; LoRA adapter weights stay in bfloat16. These are separate concerns.
- **Chat template
[Description truncada. Veja o README completo no GitHub.]