The HowToCaption dataset provides high‑quality textual captions generated from the HowTo100M instructional video dataset.It contains no video files — only captions for video clips generated using large language models (LLMs), as described in the accompanying ECCV 2024 paper.
A typical structure for this dataset on HuggingFace looks like:
HowToCaption/
│
├── metadata/
│ ├── howtocaption_train.parquet
│ ├──… See the full description on the dataset page:
https://huggingface.co/datasets/CVML-TueAI/HowToCaption.
