Real-world human–assistant logs reveal the edge cases and distributional shifts that synthetic or crowd-sourced instruction data miss. WildChat-4.8M consolidates 3,199,860 filtered conversations collected from ChatGPT interactions and exposes the metadata and moderation signals needed to study real user behavior and fine‑tune instruction-following models.
What Sets It Apart
- High scale with provenance: contains ~3.2M conversations (train split) with per-turn timestamps, turn identifiers, and usage token counts, enabling sequence-level and cost analyses rather than only isolated prompts.
- Rich contextual metadata: request headers, hashed IP, inferred state/country, and system fingerprints let researchers analyze client diversity, language/locale effects, and session linkage while preserving de‑identified traces.
- Safety-aware filtering and annotations: this release intentionally excludes conversations flagged as toxic by OpenAI Moderation or Detoxify; it also stores per-utterance moderation scores and multiple toxicity metrics for downstream safety evaluation and selective sampling.
- Reproducible pipeline: the authors released data processing scripts (including TruffleHog scanning for secrets removal) and documented the data cutoff (up to, but excluding, 2025-08-01), supporting auditability and dataset reuse.
Who It's For and Trade-offs
Great fit if you need realistic user queries and assistant responses to (1) instruction‑fine‑tune LLMs, (2) benchmark model behavior on real-world requests (WildBench subset linkage), or (3) study moderation, multilingual usage, and session dynamics. Look elsewhere if you require raw, unfiltered conversations containing toxic content (the full gated WildChat-4.8M-Full exists for approved research) or if strict privacy-compliant datasets with no hashed IP/geo are mandatory for your organization. Note also that this release filters out toxic turns, which reduces exposure to adversarial or harmful user data important for some robustness studies.
Where It Fits
Use this dataset when realism and metadata-rich analysis matter more than exhaustive exposure to harmful content. It complements synthetic instruction datasets and smaller curated conversational corpora by offering scale and operational telemetry useful for deployment-oriented research.
Practical Notes
The Hugging Face dataset card documents field schemas (conversation list, moderation structs, detoxify outputs), a train split with ~3.2M examples, and a download size/dataset size estimate. Citation pointers and contact information (e.g., Yuntian Deng) are provided on the dataset card for data‑removal requests and provenance questions.
