Visual-CoT-Key

A visual question-answering subset filtered from deepcs233/Visual-CoT. This data.parquet has 70,229 training examples. Images are not included.

Questions, boxes, and image paths come from the parent dataset. thinking and answer are written by the API model qwen3.8-flash. A local Qwen3.5-9B only checks whether the generated answer is equivalent to the reference. The local completion marker records the row count, the parent dataset, and the token range. It does not record the generator name. Before uploading, confirm from your own generation log that this parquet was produced by qwen3.8-flash.

Filtering

A candidate must satisfy both vision-token counts:

• Qwen2-VL smart_resize, with no resolution cap
• token counts for patch sizes 28 and 32 both lie in 512–4096

A generated reply is kept only when it has exactly one <answer></answer> pair, the answer is non-empty, and <key-obj> tags are balanced. A reply with no key-object tags can still be kept. It is written to parquet only after the local model judges it equivalent to the reference. The reference includes the parent's short answer and, when present, full_answer.

Fields

All six columns are strings. images, bboxs, thinking, and meta contain JSON text.

Column	Content
images	JSON list of paths relative to the parent dataset directory, for example ["images/xxxx.jpg"]. The files live under Visual-CoT images/.
bboxs	Boxes from the parent dataset, as JSON
question	Question from the parent dataset
thinking	Full qwen3.8-flash reply. The first sentence marks key objects with <key-obj>. Later paragraphs describe object state and reasoning. The last line is <answer>.
answer	Text extracted from <answer> after the equivalence check
meta	JSON object. Parent fields other than question, answer, boxes, image, reasoning, and thought are kept here. answer is the parent's original short answer. full_answer is kept when the parent has it.

The answer column is the generated answer after the check. The parent's original short answer is answer inside meta.

Images

This release does not copy or publish images. Obtain deepcs233/Visual-CoT separately, then resolve the relative paths in images against that dataset directory.

License

The parent dataset page on Hugging Face is currently marked CC BY-NC 4.0. The parent README header says apache-2.0, and the paper appendix says the data is CC BY 4.0. Those three statements disagree. This release follows the Hugging Face page and uses CC BY-NC 4.0. It does not claim a broader license.

Images remain under Visual-CoT and its upstream image sources. Qwen3.5-9B weights are Apache-2.0. Terms for the qwen3.8-flash API are those of its provider.

Citation

Please also cite Visual-CoT:

@inproceedings{shao2024visualcot,
  title={Visual CoT: Advancing Multi-Modal Language Models with a Comprehensive Dataset and Benchmark for Chain-of-Thought Reasoning},
  author={Shao, Hao and Qian, Shengju and Xiao, Han and Song, Guanglu and Zong, Zhuofan and Wang, Letian and Liu, Yu and Li, Hongsheng},
  booktitle={Advances in Neural Information Processing Systems},
  year={2024}
}