cuphead-action

# Cuphead video clips and action annotations This repository contains two source collections, kept under `rjt/` and `gh/`. Each clip consists of an MP4 video and a matching protobuf action annotation. Videos are preserved byte-for-byte from the selected source files. The RJT collection retains source-resolution videos. Some GH files were already cropped to the game area upstream. The 192/256/384/512 training resize variants, training features, model checkpoints and system backups are not included. Files appear gradually during the first upload. Check `metadata/upload_status.json` for the completion state. ## Layout ``` rjt/<video_id>/<video_id>_chunk_<number>.mp4 rjt/<video_id>/<video_id>_chunk_<number>.proto gh/<video_id>/<video_id>_chunk_<number>.mp4 gh/<video_id>/<video_id>_chunk_<number>.proto metadata/rjt.jsonl metadata/gh.jsonl metadata/summary.json ``` RJT has 5,026 clips; GH has 6,492 clips. There is no overlap in source video IDs or chunk IDs between these video collections. Full content-level video deduplication has not been asserted. Metadata records preserve each file's source path relative to the external disk and note recovered source files. One empty GH source video is supplied by its nonempty processed counterpart. ## Reading the data Download files with the ModelScope Hub SDK, then read `metadata/rjt.jsonl` or `metadata/gh.jsonl`. Each record has `video` and `annotation` paths relative to the repository root. The manifest does not define a train/validation/test split. The `.proto` files are serialized Open Pixel2Play `VideoAnnotation` messages, including per-frame gamepad actions and source metadata. This is a source-data upload: existing frame-count discrepancies and upstream game labels, including sources labeled `other`, are preserved. ## Sources and terms The action annotations originate from the [NVIDIA NitroGen dataset](https://huggingface.co/datasets/nvidia/NitroGen), whose upstream card specifies CC BY-NC 4.0. Gameplay video rights and applicable terms remain with their original sources and rights holders. This collection does not grant additional rights to the videos. Consult the upstream datasets and source metadata before reuse.