ShaoRun/vscaffold-4b-path-tracing

VScaffold-4B · Path Tracing

A task-specific checkpoint for Visual Scaffolding. The model requests visual evidence during reasoning and uses retrieved image features to continue its trace. This checkpoint handles Path Tracing in NMR-Bench. The Burden Counting checkpoint is released separately.

📄 Paper: Visual Scaffolding: Toward Native Multimodal Reasoning

Built on Qwen3-VL-4B-Thinking, the repository contains a complete model checkpoint, processor, tokenizer, and Visual Scaffolding configuration.

🚀 Run on NMR-Bench

Use Python 3.12 and a CUDA-compatible PyTorch installation.

git clone https://github.com/GeoX-Lab/VisualScaffolding.git
cd VisualScaffolding
pip install -e '.[inference]'
pip install huggingface_hub
hf download ShaoRun/nmr-bench --repo-type dataset --local-dir benchmark
hf download ShaoRun/vscaffold-4b-path-tracing --local-dir checkpoints/path_tracing
python -m visual_scaffolding evaluate \
  --task path_tracing \
  --benchmark benchmark \
  --checkpoint checkpoints/path_tracing \
  --device cuda:0 \
  --output outputs/path.jsonl

The evaluator writes predictions and a scoring summary. For single-image inference, see the GitHub repository.

📄 License

Released under Apache 2.0. See LICENSE and NOTICE.md. Checkpoint details are in model_info.json.