Test-Time Dreaming Graph Co-Evolution 9B
This is the final Qwen3.5-9B checkpoint for the second paper's trajectory SFT run (1,176 optimizer steps and 3 epochs). The dataset contained 100,116 call-level examples, trained with a 32,768-token limit and an effective batch size of 256. The loss was applied only to the final assistant response, including reasoning and answer. The vision tower and multimodal projector were frozen.
The repository contains full BF16 weights in Safetensors format; no adapter merge is required. Install the dependencies in requirements-inference.txt and run inference.py with a CUDA-enabled PyTorch environment. The upstream image/video processor settings are retained, but multimodal generation quality was not evaluated for this SFT run.