WildChat Real Human-ChatGPT Conversations

## Overview WildChat-1M is a corpus of approximately 1 million real-world conversations between human users and OpenAI's ChatGPT (GPT-3.5 and GPT-4), collected by AI2 by offering free ChatGPT access in exchange for consented chat logging. Each row is a full multi-turn conversation paired with metadata including hashed IP, country, state, request headers, timestamps, detected language, model used, and toxicity classifications. Distributed as Parquet. Time coverage runs from April 2023 through mid-2024. ## Schema - `conversation_hash` — string — unique conversation ID - `model` — string — ChatGPT model variant (gpt-3.5-turbo, gpt-4, etc.) - `timestamp` — datetime — when the conversation occurred - `conversation` — list[dict] — full turn-by-turn messages with role/content - `turn` — int — number of user-assistant turns - `language` — string — detected primary language - `openai_moderation` — list[dict] — per-turn OpenAI moderation API scores - `detoxify_moderation` — list[dict] — per-turn Detoxify toxicity scores - `toxic` — bool — aggregate toxicity flag - `redacted` — bool — whether PII redaction was applied - `state`, `country`, `hashed_ip`, `header` — user/request demographics ## Sources - allenai/WildChat-1M on HuggingFace — https://huggingface.co/datasets/allenai/WildChat-1M — License: ODC-BY - Paper: Zhao et al., "WildChat: 1M ChatGPT Interaction Logs in the Wild" (ICLR 2024) — https://arxiv.org/abs/2405.01470 ## Methodology AI2 deployed public chatbot interfaces (HuggingFace Spaces) offering free access to GPT-3.5 and GPT-4 in exchange for users consenting to have their conversations logged. Collection ran from April 2023 onward. Conversations were retained verbatim along with request metadata (IP hashed for privacy, headers, geographic info from IP geolocation). PII redaction was applied to detected emails, phone numbers, and similar identifiers. Toxicity was scored post-hoc using both the OpenAI Moderation API and Detoxify. Language was auto-detected per conversation. ## Known gaps & limitations - User population is self-selected (those seeking free GPT-4 access) and skews toward technical and non-English-speaking users versus the general ChatGPT userbase. - Geographic metadata derives from IP geolocation and may be inaccurate for VPN users. - Contains toxic, unsafe, and adult content — the dataset is intentionally unfiltered. Toxicity flags are provided but not removed. - PII redaction is best-effort and not guaranteed comprehensive; downstream users handling the data should apply additional filtering. - Conversations reflect the behavior of specific ChatGPT snapshots from 2023-2024 and may not represent current model behavior. - Multilingual but English and Chinese dominate; long-tail language coverage is sparse. ## Intended use & out-of-scope - IS for: instruction-tuning data, studying real-world user-LLM interaction patterns, alignment and safety research, multilingual chat analysis, RAG over real prompt distributions, toxicity and jailbreak research. - NOT for: deployment without safety filtering (contains toxic content), demographic inference on individuals (IPs are hashed for a reason), or training models intended to be evaluated on benchmarks that may overlap with conversation content — no dedup against common eval suites. _Federated dataset: 14 parquet shards, 3.13 GB total. Queries and downloads stream through the DataBazaar API._ Original supplier listing: WildChat-1M: 1 Million Real Human-ChatGPT Conversations 1M real-world conversations between human users and ChatGPT (GPT-3.5/4), with demographics, timestamps, languages, and toxicity labels. Parquet format, ODC-BY licensed. Widely used for instruction tuning, eval, and alignment research.