medical-sft-lora

<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->

medical-sft-lora

This model is a fine-tuned version of unsloth/llama-3-8b-bnb-4bit on the FreedomIntelligence/medical-o1-reasoning-SFT dataset, a specialized dataset designed for medical question-answering with chain-of-thought reasoning.

Model description

This model has been supervised fine-tuned using parameter-efficient fine-tuning (PEFT) with LoRA adapters on a medical question-answering dataset. It is intended to offer detailed, clinically reasoned answers to medical queries by following explicit chain-of-thought explanations, improving on the base LLM’s understanding of medical reasoning and diagnosis.

Intended uses & limitations

• Intended for research and educational purposes involving medical question-answering and reasoning.
• Useful for generating detailed step-by-step medical explanations and diagnoses.
• Not intended to replace professional medical advice, diagnosis, or treatment.
• Users should critically assess outputs and consult qualified healthcare professionals for decision-making.

Training and evaluation data

• Dataset: FreedomIntelligence/medical-o1-reasoning-SFT (English split).
• Dataset samples contain medical questions, chain-of-thought detailed reasoning, and final concise answers.
• The first 500 samples were used for supervised fine-tuning according to assignment requirements.

Training procedure

Fine-tuning was performed on the base model "unsloth/llama-3-8b-bnb-4bit", a 4-bit quantized LLM suitable for Colab GPU environments. The training used Hugging Face Transformers and PEFT libraries with the following steps:

• Prompt templates were engineered to enforce detailed clinical reasoning and stepwise explanations.
• Each training example combined a medical question, chain-of-thought reasoning, and final answer into the prompt-response pair.
• LoRA adapters were applied, targeting the q_proj and v_proj transformer modules.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 1
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 8
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• num_epochs: 3
• mixed_precision_training: Native AMP (fp16)

Training results

• Post-fine-tuning evaluation showed a marked improvement over the base model.
• The fine-tuned version generates more accurate medical reasoning with logical step-by-step explanations aligned with clinical knowledge.
• Outputs included clearer diagnoses and more relevant responses as compared to the base model.

Framework versions

• PEFT 0.17.1
• Transformers 4.55.4
• Pytorch 2.8.0+cu126
• Datasets 3.6.0
• Tokenizers 0.21.4