ch1415926/cuphead-yolo

Cuphead Qwen Teacher YOLO Dataset v2

Private research dataset for training a three-class local detector that will replace Qwen API annotation over the remaining Cuphead/UniVAM frames.

Because the images are frames from a commercial game, keep this repository private unless the project owner has independently verified redistribution rights. The license: other metadata is intentionally conservative and does not grant rights to the underlying game artwork.

Contents

code/                         portable training/evaluation/inference package
data/yolo_teacher_v2/         directly trainable YOLO dataset
provenance/final_teacher/     accepted Qwen records and quality decisions
UPLOAD_INSTRUCTIONS.md        upload and target-machine download commands

Dataset statistics:

• 2,500 images: 2,100 train and 400 validation.
• 8,512 boxes: 1,935 player, 3,609 npc, 2,968 attack_object.
• 460 explicit empty-label images, including 455 non-gameplay negatives.
• No invalid boxes and no recording-source leakage between train and validation.

Class mapping:

0 player
1 npc
2 attack_object

The validation split is held out by top-level recording source, rather than random individual frames, to avoid near-duplicate source leakage.

Train

After downloading the repository on the 80GB GPU machine:

cd code
PYTHON_BIN=python3.10 ./setup_env.sh .venv
source .venv/bin/activate

BATCH_SIZE=32 ./run_training_pipeline.sh \
  ../data/yolo_teacher_v2/data.yaml 0

The official yolo26s.pt checkpoint is downloaded by the pipeline and verified by SHA256. It is deliberately not redistributed in this dataset repository.

Do not use batch=-1 on an 80GB GPU for this small dataset. It targets roughly 60% GPU memory and can make the batch so large that each epoch has too few optimizer updates. Start with batch 32; batch 48 is a secondary experiment.

Provenance

The YOLO export was generated from 2,500 accepted Qwen teacher records. Non-gameplay records use an explicit empty annotation. The source JSONL and the final quality manifest are retained under provenance/ so the export can be audited.