Qwen Medical QLoRA Adapter

This repository contains LoRA adapter weights only. It does not contain merged or full base-model weights.

Base model: Qwen/Qwen3.6-27B

Training mixture:

json
Copy code
{
  "source_counts": {
    "medmcqa": {
      "train": 17500,
      "eval": 70,
      "weight": 0.35
    },
    "pubmedqa": {
      "train": 15000,
      "eval": 60,
      "weight": 0.3
    },
    "medqa_usmle": {
      "train": 10000,
      "eval": 40,
      "weight": 0.2
    },
    "mednli": {
      "train": 5000,
      "eval": 20,
      "weight": 0.1
    },
    "medical_safety": {
      "train": 2500,
      "eval": 10,
      "weight": 0.05
    }
  },
  "weights": {
    "medmcqa": 0.35,
    "pubmedqa": 0.3,
    "medqa_usmle": 0.2,
    "mednli": 0.1,
    "medical_safety": 0.05
  },
  "train_rows": 50000,
  "eval_rows": 200
}

Safety scope: medical education and biomedical research reasoning, not direct patient care.