DeepRadiology/Qwen3.5-9B-CXR-v3b-ckpt250-GGUF

Qwen3.5-9B CXR Structured Reporter — v3b checkpoint-250 (Q4KM imatrix GGUF)

Private research checkpoint. A QLoRA supervised fine-tune of Qwen/Qwen3.5-9B (multimodal) that turns a frontal chest radiograph — optionally with a prior study for comparison — into a six-field structured report:

Support Devices: ...
Cardiomediastinum: ...
Lungs: ...
Pleura: ...
Musculoskeletal: ...
Upper Abdomen: ...

This is the v3b training run, checkpoint-250, which was the strongest checkpoint on support-device identification in internal evaluation.

Files

Both files must be in the same folder. The text GGUF alone loads and chats but cannot see images without the mmproj.

Quantization

• —Q4_K_M guided by an importance matrix (imatrix) computed on this checkpoint's own eval generations (sentence-case structured reports — the exact surface form it emits), so the tokens that matter for radiology phrasing are protected during quantization.
• —output.weight and token_embd.weight kept at q8_0 to protect the language head (prevents dropped-letter corruption seen with vanilla Q4 on this model family).

Usage

LM Studio

1. 1.Place both .gguf files in one model folder.
2. 2.Settings → Runtime: update the llama.cpp/GGML runtime to a build recent enough to know qwen35 and its Qwen3-VL vision, or images will be ignored.
3. 3.Load, attach a frontal CXR, and prompt for a structured report.

llama.cpp (multimodal CLI)

llama-mtmd-cli \
  -m qwen35-9b-cxr-v3b-ckpt250-Q4_K_M-imat.gguf \
  -mm mmproj-qwen35-9b-cxr-v3b-ckpt250-f16.gguf \
  --image chest_xray.png \
  -sys "You are an expert chest radiologist." \
  -p "Write a structured report for this chest X-ray with sections for support devices, cardiomediastinum, lungs, pleura, musculoskeletal, and upper abdomen." \
  --temp 0 -n 400 -ngl 99

Intended use & limitations

Research artifact for chest-radiograph report-generation experiments. Not a medical device and not for clinical use. Outputs can contain errors (missed findings, hallucinated devices) and must not inform patient care. Trained on frontal views; behavior on non-frontal or non-CXR images is undefined.