UltraChat 200k
Hugging Face H4 team (filtered from Tsinghua's UltraChat)
A 200,000-dialogue filtered slice of ChatGPT-generated conversations that trained Zephyr-7B.
A heavily filtered version of UltraChat, which has 1.4 million dialogues generated by ChatGPT. Filtering fixed casing, removed refusals, and dropped low-quality turns. One row is one multi-turn conversation split into SFT and generation sets.
Access
- Cost
- FreeNo charge.
- License
- Open licenseMIT
- Sign-up
- No sign-up
- How you get it
- API and Bulk downloadParquet
What's in it
- One row is
- One row per multi-turn dialogue
- Covers
- Not geographic
- Years
- 2023
- Updated
- One-time release
- Size
- Roughly 200,000 dialogues (100K to 1M rows)
Quality
Useful baseline for chat fine-tuning, but all dialogues, including the user side, are model-written.
- Both user and assistant turns are ChatGPT output
- OpenAI terms of use may restrict training competing models
- Style is homogeneous
Connects to
- UltraFeedback: Preference set built from prompts that include UltraChat
- OpenAssistant Conversations (OASST1): Human-written alternative
Questions it could answer
- How diverse are first-turn topics compared with human-written sets?
- What share of dialogues are single-topic follow-ups?