UltraChat Multi-Round Dialogue Dataset

## Overview UltraChat is a large-scale, open-source multi-round dialogue dataset containing over 1.5 million high-quality conversations in English, generated by OpenBMB using two ChatGPT Turbo APIs in a role-play setup (one playing the user, one the assistant). The dataset is distributed in JSON format and is widely used as instruction-tuning data for chat-oriented LLMs (notably the UltraLM and Zephyr model families). ## Schema - `id` — string — unique conversation identifier - `data` — list[string] — alternating user/assistant turns within a multi-round dialogue The schema is intentionally minimal: each row is one multi-turn conversation represented as an ordered list of utterances. ## Sources - HuggingFace: https://huggingface.co/datasets/openbmb/UltraChat — license: MIT - Paper: "Enhancing Chat Language Models by Scaling High-quality Instructional Conversations" (arXiv:2305.14233) - GitHub: https://github.com/thunlp/UltraChat ## Methodology Per the source, UltraChat was generated synthetically using two ChatGPT Turbo APIs in a dual-agent setup: one model is prompted to act as a user generating realistic queries, while the other generates assistant responses. To address privacy and copyright concerns, the authors deliberately avoided seeding prompts directly from internet-scraped data. Topics were structured across three sectors: Questions about the World, Writing and Creation, and Assistance on Existing Materials. Conversations span multiple turns with the user model instructed to follow up naturally. ## Known gaps & limitations - Fully synthetic — inherits biases, hallucinations, and stylistic tics of ChatGPT (GPT-3.5 Turbo circa 2023); not a substitute for human conversation data. - English-only; not suitable for multilingual training without augmentation. - May contain factual errors since assistant responses are model-generated and not human-verified. - Potential overlap with public benchmarks — not deduplicated against common eval suites. - Source does not exhaustively document topic distribution or safety filtering coverage; buyers should validate empirically. ## Intended use & out-of-scope - **Intended:** Instruction tuning, SFT for chat assistants, dialogue modeling research, distillation targets, RAG-style conversational priors. - **Out-of-scope:** Factual grounding/QA without verification, benchmark training (leakage risk), production safety-critical applications without additional filtering, multilingual fine-tuning. _Federated dataset: 10 parquet shards, 2.34 GB total. Queries and downloads stream through the DataBazaar API._ Original supplier listing: UltraChat — Large-Scale Multi-Round Dialogue Dataset Open-source large-scale multi-round dialogue dataset (1.5M+ conversations) generated via dual ChatGPT Turbo API role-play. Widely used for instruction tuning and chat model fine-tuning. MIT licensed.