ARIVE

ARIVE adapts Qwen3-VL-4B-Instruct for grounded chest X-ray report generation. The model identifies anatomical regions, produces bounding boxes, and uses a local vision branch to inspect image crops before continuing generation.

This repository contains the final trained checkpoint in BF16: the decoder, global vision branch, and local vision branch. It was fine-tuned using SFT and GRPO on MIMIC-CXR-derived data. Training and inference prompts use Indication; additional clinical fields are used during data construction.

Inference

Inference requires the accompanying ARIVE code repository and its dependencies. The local vision branch and iterative crop observations require the project's custom inference entry point.

Download this model repository to a local directory named ARIVE and obtain the Qwen3-VL-4B-Instruct base model separately. From the ARIVE code repository root, configure the model paths and export the decoder weights for vLLM:

export QWENVL_INFER_BASE_MODEL=/path/to/Qwen3-VL-4B-Instruct
export QWENVL_INFER_VISION_CHECKPOINT=/path/to/ARIVE
export QWENVL_INFER_VLLM_CHECKPOINT=/path/to/ARIVE-vllm

python tools/export_vllm_checkpoint.py \
  --src "$QWENVL_INFER_VISION_CHECKPOINT" \
  --dst "$QWENVL_INFER_VLLM_CHECKPOINT"

The exporter first creates an ARIVE-safe intermediate directory that retains both vision branches, then creates ARIVE-vllm for decoding. Both output directories must be absent or empty. Keep the tokenizer, processor, and chat_template.jinja files with the checkpoint; the exporter copies them.

Configure the evaluation reports, reference reports, clinical context, and image paths as described in the code repository's README, then run:

CUDA_VISIBLE_DEVICES=0 python inference/infer_mcot.py

Intended use and limitations

ARIVE is intended for research on chest X-ray report generation and visual grounding. It is not intended for clinical diagnosis or patient care. Generated findings and bounding boxes may be inaccurate and require independent review.

License and data

The model checkpoint is released under Apache-2.0. See NOTICE for the base model attribution and a description of the modifications.

MIMIC-CXR images, reports, and patient-level training annotations are not included. Access to and use of the source data remain subject to the applicable PhysioNet license and data use agreement. The model license does not grant access to or permission to redistribute those datasets. The intended-use statement above describes the model's intended application and does not add restrictions to Apache-2.0.