This submission uses CATA-Qwen3B, a curriculum-aware structural VQA system that combines a frozen ViT visual encoder, online lesion-prior/topological feature extraction, full structural fusion, and Qwen2.5-3B-Instruct with QLoRA r16 plus deep TopoAdapter layers. The model was trained with curriculum alignment warmup and gate regularization before full-data continuation.

CATA-Qwen3B extends the full structural TopoAdapter architecture with a curriculum-aware training recipe. The first stage warms up language adaptation, while later training restores spatial/topological alignment losses and gate regularization. At inference time, the extracted structural features condition both visual fusion and the Qwen decoder TopoAdapters.

Install the required packages with:

The script also loads the base LLM from Hugging Face:

From this folder, run:

The system will load the checkpoint from:

It will run prediction on the Task 1 test split and write the output file: