Each Twitch channel is stored as its own dataset config, with its chat messages under data/
/. Data arrives as small append-only chunk files (data//.jsonl) added on every publish cycle -- existing chunks are immutable and never re-uploaded, so cost per publish only scales with new messages, not the dataset's total size. Sharding chunks into a per-channel directory also keeps any single directory well under the Hub's 10… See the full description on the dataset page: https://huggingface.co/datasets/Daankular/twitch-chat.
