Real user interactions — not synthetic benchmarks — reveal how people prompt LLMs, where models fail, and what safety gaps appear in the wild. This dataset captures that signal at scale, providing concrete, diverse examples for analysis, moderation model development, and instruction-following training.
What Sets It Apart
- Scale and diversity: 1,000,000 conversations collected from ~210K unique IPs across 154 detected languages, drawn from three web interfaces (single-model, side-by-side, and battle) during Apr–Aug 2023. This breadth highlights real-world multilingual and multi-turn usage patterns.
- Multi-model coverage: conversations involve 25 different state-of-the-art models (open-source and proprietary), enabling cross-model comparison and model selection studies rather than single-model bias.
- Rich per-message metadata: each record includes conversation ID, model name, conversation text in OpenAI-style JSON, detected language tag, and OpenAI moderation API labels — facilitating safety benchmarking and automated filtering experiments.
- Practical stats included: average turns ≈ 2.0, avg tokens per prompt ≈ 69.5, avg tokens per response ≈ 214.5 — useful baselines for prompt engineering and cost estimation.
Who It's For & Trade-offs
Great fit if you want to:
- Analyze real-world prompt distributions and multilingual usage patterns.
- Train or benchmark content-moderation models using real unsafe examples paired with moderation outputs.
- Build or fine-tune instruction-following/chat models with realistic conversational data or craft challenging benchmark questions.
Look elsewhere if you need:
- Fully decontaminated academic-benchmark-free corpora: the dataset did not undergo decontamination and may contain popular benchmark/test questions.
- Guaranteed removal of all PII: the curators made best-effort PII removal but retained original conversations (including unsafe content) to support safety research.
- A permissive license explicitly attached in the dataset metadata: license information is not provided alongside the dataset card and should be checked before downstream use.
Practical notes
- Collected on the Vicuna demo and Chatbot Arena web interfaces (Apr–Aug 2023) with user consent via site terms of use.
- Includes unsafe/toxic conversations intentionally retained and annotated with OpenAI moderation outputs to enable realistic safety research and to study moderation behavior.
- Data format: parquet records with conversation JSON structures; common tooling (datasets/dask/polars) can be used to process it at scale.
