Medcat V10 Weights

This repository contains the model assets for the public Medcat V10 four-task inference release. The executable inference source is maintained at HuiLin0220/Medcat.

Built with Qwen. The bundle contains the InternVL3-8B-hf base model, task/source LoRA adapters, and compact classification heads required for disease classification, multi-label classification, detection, and regression.

Download Into the Source Repository

Clone the inference source, install the Hugging Face CLI, and download only the model tree into the repository root:

git clone https://github.com/HuiLin0220/Medcat.git
cd Medcat
python -m pip install --upgrade huggingface_hub
hf download HuiLin0220/Medcat --include "models/**" --local-dir .
sha256sum --check SHA256SUMS

The resulting layout is:

Medcat/
|-- inference.py
|-- predict.sh
`-- models/
    |-- InternVL3-8B-hf/
    `-- FLARE-InternVL3-8B-hf/

The model directory is approximately 16 GB. Downloads use the Hugging Face Hub's resumable transfer mechanism.

Integrity

SHA256SUMS.models records every file in the model tree. From a checkout of this weight repository, verify it with:

sha256sum --check SHA256SUMS.models

All tensor files were compared byte-for-byte with the evaluated Medcat V10 container. One unused predecessor focal checkpoint was omitted, and two machine-local provenance paths in non-tensor metadata were replaced with portable descriptions.

Licenses

• Medcat and ME-VLIP components: LICENSE-MEDCAT (Apache-2.0)
• InternVL components: LICENSE-INTERNVL (MIT)
• Qwen components: LICENSE-QWEN (Qwen License Agreement)
• Required attribution notices: NOTICE

Users must comply with all applicable upstream terms. These weights are released for research and challenge reproduction and are not intended for clinical diagnosis or treatment.