qwen2.5-7b-medmcqa-qlora-5k

Qwen2.5-7B-Instruct — MedMCQA QLoRA adapter (5K)

A 4-bit QLoRA adapter for Qwen/Qwen2.5-7B-Instruct, fine-tuned on MedMCQA as one point of a data-scaling study of when parameter-efficient domain fine-tuning helps an already-strong instruction-tuned LLM. Powers the RAG assistant at https://github.com/Udit013/biomed-llm-peft (live demo: https://huggingface.co/spaces/Udit013/biomed-assistant).

⚠️ Intended use & limitation

Research and education only. Must NOT be used for real clinical decisions, diagnosis, or treatment. Outputs are uncalibrated and may be confidently wrong.

Training

• Base model: Qwen/Qwen2.5-7B-Instruct
• Method: 4-bit QLoRA (nf4, double quant), LoRA r=16, alpha=32, dropout=0.05
• Trainable params: 40,370,176 / 4,393,342,464 (0.918894%)
• Data: MedMCQA, N=5000 (seeded subsample), 1 epoch
• Seed: 42 · LR: 0.0002 · eff. batch = 1×16
• Final train loss: 0.868

Evaluation (EleutherAI lm-evaluation-harness, matched 200-item subsample)

In-domain is flat (within noise) — strong instruction tuning already near-saturates MedMCQA; the OOD gain is likely an answer-selection/format effect, not new knowledge. See the repo for the full methodology and honest write-up.

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-7B-Instruct", device_map="auto", load_in_4bit=True)
model = PeftModel.from_pretrained(base, "Udit013/qwen2.5-7b-medmcqa-qlora-5k")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct")