🧠 Clinical MCQ Solver using LoRA-tuned Language Models

This project fine-tunes a causal language model (Phi-3.5) using LoRA (Low-Rank Adaptation) to solve multiple-choice clinical reasoning tasks. It was developed as part of the Goedel Machines x IITM Clinical LLM Challenge.

🚀 Features

• Preprocessing of clinical MCQs into generative-style prompts
• Token-level loss masking to focus learning only on the answer portion
• LoRA fine-tuning with PEFT on HuggingFace Transformers
• Evaluation logic for predicting MCQ answers
• Submission pipeline generating CSV for Kaggle-style competitions

🛠️ Setup

Install dependencies:

pip install -q \
    transformers==4.40.2 \
    peft==0.11.0 \
    accelerate==0.30.0 \
    datasets

Ensure you have GPU access (cuda) enabled.

🧩 Training Pipeline

• Tokenizer & Data Prep: Converts question + options into a prompt
• Loss Masking: Loss computed only on the answer text
• Model: Phi-3.5 Mini Instruct + LoRA (targeting qkv_proj, o_proj)
• Trainer: HuggingFace Trainer API with gradient_checkpointing + FP16

Run training:

trainer.train()

📈 Evaluation Logic

Evaluation decodes model output and compares predicted vs. true answers based on the string that follows "Answer:".

Accuracy is computed across:

• Easy
• Medium
• Hard test sets.

📦 Submission

The submission.ipynb file:

• Loads the fine-tuned model with adapter weights
• Performs predictions on test data
• Saves results to submission.csv (id → answer)

📂 Directory Structure

.
├── medical-llm.ipynb     # Training + preprocessing notebook
├── submission.ipynb      # Evaluation + submission script
├── /lora_phi3_stage1     # (output model dir)
└── submission.csv        # Final submission

📌 Notes

• Uses LoRA with PEFT for parameter-efficient fine-tuning
• Only trains specific layers of the base model for speed & memory
• Designed to work on Kaggle Kernel environments (uses /kaggle/working)

👥 Authors

This solution was developed by Ritankar.