Fine-tuning VisualGLM to read chest X-rays in Chinese
Because medical vision-language models are mostly English-only, XrayGLM fine-tunes VisualGLM-6B on ChatGPT-translated chest X-ray reports so clinicians can query radiographs in Mandarin.

What it does
XrayGLM is essentially a fine-tuned VisualGLM-6B that ingests chest X-rays and writes diagnostic impressions in Chinese. It handles both single-shot report generation and multi-turn conversation—users can ask what the image shows, whether specific conditions like emphysema appear, or even request diet advice. The authors curated a Chinese dataset by translating public English radiology reports from OpenI and MIMIC-CXR with ChatGPT.
The interesting bit
The real labor here is dataset construction, not architecture. The team took existing public X-ray corpora, structured the messy unstructured reports, and used ChatGPT to produce Chinese diagnostic text, creating a bilingual medical vision-language dataset where none existed. The released LoRA checkpoints are explicitly labeled low-quality, and side-by-side examples show the model occasionally flatly contradicting the reference radiologist—declaring a mildly enlarged heart “normal,” for instance.
Key highlights
- Fine-tuned
VisualGLM-6Bwith LoRA; training used 4× A100 80GB GPUs - Two low-quality checkpoints (
checkpoints-XrayGLM-300and3000) on Hugging Face for academic research; a higher-quality “plus” variant is mentioned but lacks a download link - Supports both CLI and WebUI inference
- Dataset derived from
OpenI(~6,423 Chinese-translated samples) andMIMIC-CXR CC BY-NC-SA 4.0license—strictly non-commercial and academic
Caveats
- Showcased outputs are explicitly from low-quality weights, and the README warns that performance is limited
- Example comparisons reveal factual mismatches against the provided doctor reports
- High-quality weights and the full
MIMIC-CXR-zhdataset are not yet released
Verdict
A useful reference for researchers building Chinese medical VLMs or curating radiology datasets. Not remotely clinical-ready, and the authors themselves note that a 6B model will inevitably lose to larger-parameter competitors.
Frequently asked
- What is WangRongsheng/XrayGLM?
- Because medical vision-language models are mostly English-only, XrayGLM fine-tunes VisualGLM-6B on ChatGPT-translated chest X-ray reports so clinicians can query radiographs in Mandarin.
- Is XrayGLM open source?
- Yes — WangRongsheng/XrayGLM is an open-source project tracked on heatdrop.
- What language is XrayGLM written in?
- WangRongsheng/XrayGLM is primarily written in Python.
- How popular is XrayGLM?
- WangRongsheng/XrayGLM has 1.1k stars on GitHub.
- Where can I find XrayGLM?
- WangRongsheng/XrayGLM is on GitHub at https://github.com/WangRongsheng/XrayGLM.