qwen3-4b-radiology-report-lora

LoRA adapter for Qwen/Qwen3-4B-Instruct-2507, fine-tuned to edit a normal radiology report template into a final report given a radiologist's dictation - outputting a JSON diff of only the fields that changed, which a separate deterministic renderer merges back into the full report.

• Base model: Qwen/Qwen3-4B-Instruct-2507 (Apache 2.0)
• Method: QLoRA (4-bit NF4 base, LoRA r=32, alpha=64, target modules: q/k/v/o_proj, gate/up/down_proj)
• Task: private hiring-challenge dataset (radiology template editing) - trained data is not included or redistributed here, only the resulting adapter weights.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel
import torch

base_id = "Qwen/Qwen3-4B-Instruct-2507"
tokenizer = AutoTokenizer.from_pretrained("surenoobsterIITBHU/qwen3-4b-radiology-report-lora")
base = AutoModelForCausalLM.from_pretrained(
    base_id,
    quantization_config=BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4",
                                            bnb_4bit_compute_dtype=torch.bfloat16),
    device_map="auto",
)
model = PeftModel.from_pretrained(base, "surenoobsterIITBHU/qwen3-4b-radiology-report-lora")

Expects a prompt describing the template's fields, retrieved similar prior cases, and a new dictation, and returns a strict JSON object: {"changed_fields": {...}, "impression": "...", "preamble": null}.