You can caption files like this (ms-swift formatting):
{"messages": [{"role": "user", "content": "<image>Write a detailed caption for this image, maximum 1000 words. Mention pivotal elements—people, objects, scenery—using confident, definite language. Focus on concrete details like color, shape, and spatial relationships. Use vulgar descriptions of what the characters are doing."}], "images": ["josman_images/97fecadaa022a4bf7a4947edc1f3260a.webp"]}
About the model:
The model has been fine tuned on synthetic reasoning data with 500 curated high quality captions,
where the reasoning includes contrastive 'fixing its own mistakes'. It was also trained on 9000 image-tag pairs with synthetic chain of thought.
The chain of thought was generated by Gemini 3.1 Flash (tags) and Pro (captions).
All tags + captions were very high quality.
What is in this repo:
Qwen3.5 9B fused with 32 rank LoRA + fully finetuned projector (visual.merger)
