Combined Chat Datasets A standardized, unified collection of 30 conversational AI datasets spanning organic in the wild chats, voluntary sharing, side by side preferences, conversation trees, RLHF pairs, and crowdsourced instruction tuning data normalized to a single schema for easy joint use. This dataset is a re distribution. It does not relicense the underlying data. See the Legal & Licensing section you must comply with each source dataset's original license. Quick start To list all available configs: Each config corresponds to one source dataset. The schema is shared across all configs, so you can concatenate freely: Why combine these? Conversational data for LLM training and evaluation is scattered across dozens of repositories with 6 different schema patterns , inconsistent role names, varied preference encodings, and different timestamp formats. This dataset provides: 1. One unified schema every row has messages: list[{role, content}] , no matter the source. 2. One loading API load dataset(..., "config name") for all 30 datasets. 3. Per source configs load just what you need; no need to download 30+ GB to access one dataset. 4. Provenance preserved source dataset column alw…
Runs entirely in your browser via DuckDB-Wasm — this dataset's real data file is loaded once, then queried locally. Nothing is sent to a server.
We use cookies for essential functionality and analytics. You can accept or reject analytics cookies.Cookie policy