shuaishuaicdp/OmniCoding

OmniCoding是一个多模态终端工具使用数据集，专为监督微调（SFT）和强化学习（RL）而设计。每个数据记录包含一个问题、一个可验证答案和媒体（视频/音频/图像）——目标代理需要通过Linux终端（如使用ffmpeg、ffprobe、whisper、python等工具）操作媒体来回答问题，而不是通过图形用户界面。该数据集从四个上游来源（Omnimodal-Agent-SFT-2K、OmniVideoBench、AVUTBenchmark和Video-MME-v2）聚合和过滤而来，采用统一的模式、全局去重和类别平衡采样。数据集旨在训练多模态代理在终端环境中使用工具处理媒体内容。

OmniCoding is a multimodal terminal-tool-use dataset for SFT/RL. Each record consists of a question, a verifiable answer, and media (video/audio/image) — the target agent is expected to operate on the media via a Linux terminal (using tools like ffmpeg, ffprobe, whisper, python, etc.) rather than a GUI. It is aggregated and filtered from four upstream sources, with a single unified schema, global deduplication, and category-balanced sampling. The dataset is designed to train multimodal agents to use tools in terminal environments for media processing.