We open-source both the Automatically Synthesized CoM Data and the Manually Annotated CoM-Math Data to facilitate potential research. The automatically synthesized CoM data (i.e., com.jsonl) consists of 84K positive reasoning chains, which was produced by an automated data generation pipeline with an LLM-based (GPT-4) linguistic solving steps generation and a VFMs-based (GroundingDINO, PaddleOCR) visual evidence compensation upon massive public VQA samples. We… See the full description on the dataset page:
https://huggingface.co/datasets/qijimrc/CoMDataset.
