LlamaTron-RS1-Rolex

Fine-tuned version of Meta's Llama-3.2-1B-Instruct on the ReasonMed dataset (370K high-quality medical reasoning examples) using LoRA. The model naturally exhibits clear step-by-step Chain-of-Thought (CoT) reasoning on medical multiple-choice and open-ended questions.

This repository provides the merged weights and a GGUF file in FP16 format for efficient local inference.

Key Features

• Parameter-efficient fine-tuning with LoRA (~0.1–0.3% of parameters updated)
• Full support for ReasonMed chat-template conversations
• Mixed-precision training (FP16)
• Observable CoT medical reasoning
• GGUF file in FP16 format for local inference (llama.cpp, Ollama, LM Studio, etc.)

Important Disclaimer

This model is for research, education, and prototyping purposes only.
It is not a medical device, diagnostic tool, or substitute for professional clinical judgment. Always consult qualified healthcare professionals for medical decisions.

Dataset

ReasonMed – the largest publicly available medical reasoning dataset (as of 2025)

• Paper: ReasonMed: A 370K Multi-Agent Generated Dataset for Advancing Medical Reasoning
• Size: 370,000 high-quality reasoning examples
• Generation: Multi-agent LLM pipeline + Error Refiner + EMD curation
• Format: JSONL with role-based conversation turns
• License: Follow the terms set by the ReasonMed authors

Training Details

• Base model: meta-llama/Llama-3.2-1B-Instruct
• Method: LoRA (rank=8, alpha=16, dropout=0.05)
• Target modules: q_proj, k_proj, v_proj, o_proj
• Optimizer: Adafactor
• Hyperparameters:
  • Epochs: 3
  • Global batch size: 16 (per-device 4 + gradient accumulation 4)
  • Learning rate: 2e-4
  • Warmup steps: 20
  • Max sequence length: 512
• Hardware: NVIDIA H100 (rented via JarvisLabs.ai)

Post-Training Steps

1. Merged LoRA adapters into base model
2. Converted to GGUF (FP16)

Files in this Repository

• llama3.2-1b-medical-reasonmed-fp16.gguf

Note: This is the FP16 GGUF file. Users can further quantize it locally using llama.cpp (e.g., to Q4_K_M, Q5_K_M, or Q8_0) for smaller file sizes and faster inference on lower-end hardware.

Inference Example (llama.cpp)

./llama.cpp/main \
  -m llama3.2-1b-medical-reasonmed-fp16.gguf \
  --color --temp 0.7 --top-p 0.9 \
  -p "A patient presents with fever, cough, and shortness of breath. What is the most appropriate initial investigation?\nA. ECG\nB. Chest X-ray\nC. Blood culture\nD. CT pulmonary angiogram"
Limitations

1B-parameter model → best for lightweight / edge use cases
Reasoning quality lags behind larger (7B–70B) medical models
No additional instruction-tuning or preference optimization (DPO/ORPO) yet

Future Work

DPO / ORPO alignment
Fine-tuning on larger bases (Llama-3.2-3B, Meditron, etc.)
Formal evaluation on MedQA, PubMedQA, MMLU-clinical

License

Code (training/merging scripts): MIT (see GitHub repo)
Base model: Meta Llama 3.2 Community License
Fine-tuned weights & GGUF files: Same as base model + ReasonMed dataset terms

Acknowledgments

ReasonMed authors (Yu Sun et al.)
Meta AI for Llama-3.2
JarvisLabs.ai for affordable H100 access
Hugging Face, PEFT, and llama.cpp contributors

For questions or collaboration, open an issue on the linked GitHub repository or reach out via LinkedIn.