Datasetpublic

OpenMOSS-Team/GameQA-text

Here we provide a pure-text version of GameQA, encompassing some appropriate games. (See https://github.com/tongjingqi/Code2Logic/issues/2) Code2Logic: Game-Code-Driven Data Synthesis for Enhancing VLMs General Reasoning This is the first work, to the best of our knowledge, that leverages game code to synthesize multimodal reasoning data for training VLMs. Furthermore, when trained with a GRPO strategy solely on GameQA (synthesized via our proposed Code2Logic approach), multiple… See the full description on the dataset page: https://huggingface.co/datasets/OpenMOSS-Team/GameQA-text.

sourceHugging Facemitupdated 1y agoView on Hugging Face

3likes23downloads

Dataset Card

*Here we provide a pure-text version of GameQA, encompassing some appropriate games.* (See https://github.com/tongjingqi/Code2Logic/issues/2)

Code2Logic: Game-Code-Driven Data Synthesis for Enhancing VLMs General Reasoning

This is the first work, to the best of our knowledge, that leverages *game code to synthesize multimodal reasoning data for training VLMs. Furthermore, when trained with a GRPO strategy solely on GameQA (synthesized via our proposed Code2Logic* approach), multiple cutting-edge open-source models exhibit significantly enhanced out-of-domain generalization.

[📖 Paper] [💻 Code] 🤗 [GameQA-140K Dataset] 🤗 [GameQA-InternVL3-8B ] 🤗 [GameQA-Qwen2.5-VL-7B] 🤗 [GameQA-LLaVA-OV-7B ]