UltraChat 200k Zephyr SFT Dataset

## Overview UltraChat 200k is a heavily filtered subset of the original 1.4M-dialogue UltraChat corpus, curated by HuggingFaceH4 and used to train the Zephyr-7B-β chat model. It contains roughly 200,000 multi-turn dialogues generated by ChatGPT spanning a wide range of topics. The data is provided in parquet format with train/test splits for supervised fine-tuning (SFT) and generation evaluation. ## Schema - `prompt` — string — the initial user prompt that seeded the dialogue - `prompt_id` — string — stable hash identifier for the prompt - `messages` — list[dict] — multi-turn conversation as `{role, content}` pairs (roles: user/assistant) ## Sources - HuggingFace: https://huggingface.co/datasets/HuggingFaceH4/ultrachat_200k — License: MIT - Original UltraChat paper: arXiv:2305.14233 ## Methodology The original UltraChat dataset was generated by having ChatGPT play both user and assistant roles across diverse topical seeds. HuggingFaceH4 produced the 200k variant by: (1) subsetting for faster SFT iteration, (2) truecasing approximately 5% of dialogues that were observed to be incorrectly lowercased, (3) removing dialogues where the assistant refused or hedged inappropriately (e.g., "I do not have personal experiences"), and (4) further filtering for grammatical and content quality. Splits are provided for both SFT training and generation-style evaluation. ## Known gaps & limitations All content is synthetic, generated by ChatGPT — it inherits OpenAI model biases, hallucinations, and stylistic tics from that era (early-to-mid 2023). English only. The dataset is not deduplicated against common public benchmarks. Some residual refusals and assistant-style artifacts may remain despite filtering. Truecasing was heuristic and may have introduced minor errors. ## Intended use & out-of-scope - IS for: supervised fine-tuning of base LLMs into chat/instruct models, replicating Zephyr-style training, multi-turn dialogue research. - NOT for: factual QA ground truth (synthetic content, not verified), training models intended to claim non-AI persona, or as an evaluation set against models that were trained on it (leakage risk). _Federated dataset: 8 parquet shards, 1.51 GB total. Queries and downloads stream through the DataBazaar API._ Original supplier listing: UltraChat 200k (Zephyr SFT Dataset) Filtered 200k-dialogue subset of UltraChat used to train Zephyr-7B-β. High-quality multi-turn ChatGPT-generated conversations for supervised fine-tuning of chat models. MIT licensed, parquet format.