ramaalhamidi/videomme-carve

VideoMME + CARVE bundle

Raw VideoMME benchmark data plus everything needed to run the CARVE vanilla-agent / evidence-dependence pipeline against it.

Contents

• videos/ - raw VideoMME source videos (mp4, sourced from YouTube).
• subtitles/ - VideoMME subtitle files.
• embeddings/10/large/ - precomputed LanguageBind clip embeddings (clip_duration=10s, retriever_type=large), keyed by video ID, used by the retrieval stage so it doesn't need to be recomputed.
• qa.json - full VideoMME question/answer set.
• qa_e0_pilot.json - the pilot subset used for the E0 control experiment.
• code/ - the subset of the CARVE / VideoDeepResearch codebase required to actually run the vanilla agent and the CARVE probe against this data:
  • eval.py - the agent loop, tool implementations, and CARVE probe logic.
  • video_utils.py - video clip extraction / frame caching.
  • retriever_languagebind.py - the LanguageBind-based video retriever.
  • servers_io/vllm_server_planner.py, vllm_server_qwen.py, vllm_server_temporal_grounder.py - the split client/server vLLM engine wrappers (planner / VLM / temporal-grounding models) used by eval.py's file-based IPC.

Note on video content

Videos are sourced from YouTube per the original VideoMME benchmark release. This bundle is a working copy for the CARVE research project; no rights to the underlying video content are claimed.