BoxComm-Dataset
BoxComm-Dataset is the official data release for BoxComm, a benchmark for category-aware boxing commentary generation and narration-rhythm evaluation.
Resources
- Project Page: https://gouba2333.github.io/BoxComm
- Paper: http://arxiv.org/abs/2604.04419
- Code: https://github.com/gouba2333/BoxComm
- Benchmark: https://huggingface.co/datasets/gouba2333/BoxComm
Overview
This dataset release is intended for training, analysis, and reproducible preprocessing. It contains the complete processed videos together with the released annotations and benchmark metadata.
Recommended structure:
BoxComm-Dataset/
├── train/
│ ├── videos/
│ ├── events/
│ └── asr/
├── eval/
│ ├── videos/
│ ├── events/
│ └── asr/
└── metadata/
The split convention is:
train: video id< 478eval: video id>= 478
Each event directory should contain:
- one skeleton
.pklfile - one
video_event_inference_3.jsonfile
Each ASR JSON file should contain classified_segments.
What is included
- processed match videos
- event annotations
- skeleton data
- ASR with sentence segmentation
- 3-way commentary labels
- split metadata
Intended uses
- supervised fine-tuning for commentary generation
- category-aware commentary evaluation
- narration-rhythm analysis
- multimodal sports video understanding research
Data preparation in the code repository
The official code repository provides:
scripts/prep_qwen3vl_sft_data.pyscripts/train_qwen3vl.pyscripts/infer_qwen3vl.pyscripts/eval_metrics.pyscripts/eval_streaming_cls_metrics.py
Repository: https://github.com/gouba2333/BoxComm
Licensing
The public release includes processed videos, ASR annotations, event JSON files, skeleton PKL files, and benchmark metadata for research use.
Citation
@article{wang2026boxcomm,
title={BoxComm: Benchmarking Category-Aware Commentary Generation and Narration Rhythm in Boxing},
author={Wang, Kaiwen and Zheng, Kaili and Deng, Rongrong and Shi, Yiming and Guo, Chenyi and Wu, Ji},
journal={arXiv preprint arXiv:2604.04419},
year={2026}
}