Mistral-Nemo Style (step 2000) — GGUF
Fine-tuned Mistral-Nemo-Instruct-2407 LoRA (style training) merged at checkpoint step 2000, then converted to GGUF for llama.cpp / LM Studio.
Files
| File | Quant | Size (approx) | Notes |
|---|---|---|---|
mistral-style-step2000-q5_k_m.gguf |
Q5_K_M | ~8.2 GB | Recommended quality/size |
mistral-style-step2000-q4_k_m.gguf |
Q4_K_M | ~7.0 GB | Smaller / faster |
LM Studio
- Download a
.gguffrom this repo - Import into LM Studio (or place in your models folder)
- Use the Mistral chat template if not auto-detected
Suggested sampling (starting point): temperature 1.0, top_p 0.95.
Training notes
- Base:
unsloth/mistral-nemo-instruct-2407-bnb-4bit(Mistral-Nemo-Instruct-2407) - Method: QLoRA rank 32 via Unsloth / TRL SFT
- Checkpoint: step 2000 (of 3000 configured max steps)
- Conversion: merge LoRA → HF F16 →
convert_hf_to_gguf.py→llama-quantize
License
Follow the base model license (Mistral-Nemo-Instruct-2407) and Apache-2.0 for conversion tooling.