radagent-qwen3-14b-lora

RadAgent Qwen3 14B LoRA

This repository contains the RadAgent LoRA adapter for OpenPipe/Qwen3-14B-Instruct, trained for chest CT report generation within the RadAgent pipeline. Note the source code for using RadAgent within its toolbox environment will be released in due time.

Paper

📄 RadAgent: A tool-using AI agent for stepwise interpretation of chest computed tomography

• ArXiv: https://arxiv.org/abs/2604.15231
• Hugging Face Papers: https://huggingface.co/papers/2604.15231

Important

This repository contains a PEFT/LoRA adapter, not a standalone merged model.

To use it, you must also have access to the base model:

• OpenPipe/Qwen3-14B-Instruct

This repository currently contains only the adapter weights. We will also provide the RadAgent tool code and an easy inference setup for running the full framework in due time.

Training Summary

• Base model: OpenPipe/Qwen3-14B-Instruct
• Adapter type: LoRA
• Task: causal language modeling for report generation
• LoRA rank: 16
• LoRA alpha: 32
• Target modules:
• q_proj
• k_proj
• v_proj
• o_proj
• Training algorithms: GRPO
• Training framework: art

Usage With Transformers + PEFT

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

base_model = "OpenPipe/Qwen3-14B-Instruct"
adapter_repo = "RadAgent/radagent-qwen3-14b-lora"

tokenizer = AutoTokenizer.from_pretrained(base_model, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
  base_model,
  torch_dtype=torch.bfloat16,
  device_map="auto",
  trust_remote_code=True,
)
model = PeftModel.from_pretrained(model, adapter_repo)

Usage Within RadAgent

The RadAgent framework allows loading the Hugging Face adapter directly, use:

python agents/main_ct_rate.py \
--model_name='OpenPipe/Qwen3-14B-Instruct' \
--mode='val' \
--task='report_generation' \
--batch_size=128 \
--start_id=0 \
--end_id=1000 \
--inference_model_name='RadAgent/radagent-qwen3-14b-lora' \
--agent_type='v8c' \
--do_check_summary

For using the entire framework we advise to use 8 GPUs.

Limitations

• Performance depends on the surrounding RadAgent orchestration and tool-calling setup. We will release a Docker-setup and the code to run the entire toolbox together with the trained agent in due time.
• Outputs may contain omissions, factual mistakes, or hallucinations.
• Performance outside the original workflow may differ substantially.
• This model is for research use only.

License And Usage Restrictions

This adapter is provided for research use only and is not licensed for clinical use.

Permission is granted to use, reproduce, and modify this adapter solely for non commercial research, evaluation, and academic reproducibility purposes. This repository contains only adapter weights and does not include the base model weights. Use of this adapter requires separate access to OpenPipe/Qwen3-14B-Instruct and remains subject to the license terms, usage restrictions, and access conditions of the base model.

Citation

If you use this model or the RadAgent framework in your research, please cite:

@misc{roschewitz2026radagenttoolusingaiagent,
      title={RadAgent: A tool-using AI agent for stepwise interpretation of chest computed tomography}, 
      author={Mélanie Roschewitz and Kenneth Styppa and Yitian Tao and Jiwoong Sohn and Jean-Benoit Delbrouck and Benjamin Gundersen and Nicolas Deperrois and Christian Bluethgen and Julia Vogt and Bjoern Menze and Farhad Nooralahzadeh and Michael Krauthammer and Michael Moor},
      year={2026},
      eprint={2604.15231},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2604.15231}, 
}