JoyDaJun/Medragchecker-Student-Checker

MedRAGChecker Student Checker — LoRA Adapters

This repository hosts LoRA adapters for the checker component used in the MedRAGChecker project. The checker is trained as an NLI-style verifier to classify a (evidence, claim) pair into:

• Entail
• Neutral
• Contradict

These adapters are intended for research and evaluation (e.g., ensembling multiple checkers trained with different base models and/or training recipes such as SFT vs. GRPO).

This repo contains adapters only. You must load each adapter on top of its corresponding base model.

---

Contents

Each adapter subfolder typically includes:

• adapter_config.json
• adapter_model.safetensors (or .bin)

---

Available adapters

All adapters live under the Checker/ directory:

Adapter subfolder	Base model (HF id)	Training recipe	Notes
Checker/med42-llama3-8b-sft	<PUT_BASE_MODEL_ID_HERE>	SFT
Checker/med42-llama3-8b-grpo	<PUT_BASE_MODEL_ID_HERE>	GRPO
Checker/meditron-sft	<PUT_BASE_MODEL_ID_HERE>	SFT
Checker/meditron-grpo	<PUT_BASE_MODEL_ID_HERE>	GRPO
Checker/PMC_LLaMA_13B-sft	<PUT_BASE_MODEL_ID_HERE>	SFT
Checker/qwen2-med-7b-sft	<PUT_BASE_MODEL_ID_HERE>	SFT
Checker/qwen2-med-7b-grpo	<PUT_BASE_MODEL_ID_HERE>	GRPO

How to fill the “Base model (HF id)” column

Use a valid Hugging Face Hub model id (format: org/name). Examples:

• meta-llama/Meta-Llama-3-8B-Instruct
• Qwen/Qwen2-7B-Instruct

If your base model is not available on the Hub (only stored locally), you can either:

1. upload the base model to a private Hub repo and reference that id here, or
2. keep this field as N/A (local) and document your local loading instructions.

---

Quickstart: load an adapter with PEFT

from transformers import AutoModelForSequenceClassification, AutoTokenizer
from peft import PeftModel

# 1) Choose the base model that matches the adapter you want to use
base_model_id = "<HF_BASE_MODEL_ID>"

# 2) Choose the adapter subfolder inside this repo
repo_id = "JoyDaJun/Medragchecker-Student-Checker"
subfolder = "Checker/qwen2-med-7b-sft"  # example

tokenizer = AutoTokenizer.from_pretrained(base_model_id, use_fast=True)
model = AutoModelForSequenceClassification.from_pretrained(base_model_id)
model = PeftModel.from_pretrained(model, repo_id, subfolder=subfolder)

If your checker was trained using a Causal LM head instead of a sequence classification head, replace AutoModelForSequenceClassification with AutoModelForCausalLM and use the same prompt/template as in training.

---

Ensemble usage (optional)

If you trained multiple student checkers, you can ensemble them (e.g., by weighting each checker’s class probabilities using dev-set reliability such as per-class F1). This often helps stabilize performance across Entail / Neutral / Contradict, especially under class imbalance.

---

Limitations & responsible use

• Not medical advice. Do not use for clinical decision-making.
• Outputs may reflect biases or errors from training data and teacher supervision.
• Please evaluate on your target dataset and report limitations clearly.

---

Citation

If you use these adapters, please cite your MedRAGChecker paper/project:

@article{medragchecker,
  title={MedRAGChecker: A Claim-Level Verification Framework for Biomedical RAG},
  author={...},
  year={2025}
}