HoqueMahmudul/llava-onevision-0.5b-qdora-radiology-image-caption

Document DoRA inference overhead vs QLoRA across scales

Clarify latency comparability: base quantization must match, not just GPU

Correct multi-adapter section: hot-swap verified identical (0/20); divergence was an adapter-naming confound

Label GPU per measurement; correct max_new_tokens to 256; add multi-adapter serving caveat

Replace estimated VRAM with measured peak memory and latency (n=500, A100)

Link to the medical-vlm-explainability code repository

Link to the Radiology Image Captioning collection

Update README.md

Add citation

Add ImageCLEFmedical 2026 evaluation section (metrics named); update training data and recipe details

Expand docs: install, quick start, 4-bit, batching, hardware, fine-tuning, merging, troubleshooting

Document continued fine-tuning (is_trainable=True) and warm-start caveats

Fix load snippet: load processor from base model (adapter repo lacks preprocessor_config.json)

Add model card

Add fine-tuned PEFT adapter weights

initial commit