SKU SFT-2004 · Sold by External
UltraChat 200k
Product specifications
| SKU | SFT-2004 |
|---|---|
| Data type | SFT dataset |
| Volume | 207865 rows |
| Size on disk | 1.62 GB |
| Format | Parquet (Hugging Face) |
| Access model | PUBLIC LICENSE |
| Pricing | Free · open-source license |
| Quality score | — |
| License | MIT |
A cleaned and filtered subset of Tsinghua's original UltraChat corpus, prepared by the HuggingFaceH4 team for the Zephyr recipe. It retains multi-turn assistant-style dialogues covering world knowledge, creative writing/generation and rewriting/summarization, with truecasing and quality filters applied. Rows expose a prompt, a prompt_id hash, and a messages list of {role, content} turns; the dataset ships four splits (train/test for _sft and _gen).