Source-linked AI summary
WildChat: 1M ChatGPT Interaction Logs in the Wild
Wenting Zhao, Xiang Ren, Jack Hessel, Claire Cardie, Yejin Choi, Yuntian Deng
TL;DR
Public evidence of real-world chatbot use and access to instruction-tuning data remain limited. WILDCHAT collects consented, anonymized user-chatbot conversations with metadata and analyzes their breadth and toxicity. The dataset contains diverse multilingual interactions, over 10% toxic interactions, and supports instruction-tuning applications.
Problem
Public datasets of real user-chatbot interactions and instruction-tuning data remain limited because natural interactions are mostly proprietary and instruction-tuning datasets are often inaccessible.
Method
The authors deployed publicly accessible GPT-3.5-Turbo and GPT-4 chatbot services, collecting consented transcripts, request headers, and anonymized geographic metadata.
Results
WILDCHAT provides real-world multi-turn multilingual interactions, over 10% toxic interactions, and a raw dataset that produces a strong chatbot when used for fine-tuning.
Takeaways & Limitations
WILDCHAT supports research on chatbot behavior, toxicity and safeguarding, instruction-tuning, computational social science, and AI ethics.
Takeaways & Limitations
Users associated with Hugging Face Spaces and communities sharing the service may be overrepresented, while anonymity may bias the dataset toward more toxic content.
Abstract
from arXiv · showhide
Chatbots such as GPT-4 and ChatGPT are now serving millions of users. Despite their widespread use, there remains a lack of public datasets showcasing how these tools are used by a population of users in practice. To bridge this gap, we offered free access to ChatGPT for online users in exchange for their affirmative, consensual opt-in to anonymously collect their chat transcripts and request headers. From this, we compiled WildChat, a corpus of 1 million user-ChatGPT conversations, which consists of over 2.5 million interaction turns. We compare WildChat with other popular user-chatbot interaction datasets, and find that our dataset offers the most diverse user prompts, contains the largest number of languages, and presents the richest variety of potentially toxic use-cases for researchers to study. In addition to timestamped chat transcripts, we enrich the dataset with demographic data, including state, country, and hashed IP addresses, alongside request headers. This augmentation allows for more detailed analysis of user behaviors across different geographical regions and temporal dimensions. Finally, because it captures a broad range of use cases, we demonstrate the dataset's potential utility in fine-tuning instruction-following models. WildChat is released at https://wildchat.allen.ai under AI2 ImpACT Licenses.
1 INTRODUCTION
WILDCHAT addresses the limited public availability of real user-chatbot interactions and proprietary instruction-tuning data by releasing a large, consented corpus of diverse, multi-turn conversations. Its breadth supports toxicity research and instruction-tuning applications.
- Instruction-tuning datasets are often proprietary, limiting researchers’ access to data needed to advance conversational agents.
- Real user-chatbot interactions are mostly proprietary, while expert-curated datasets often differ from real-world distributions and remain single-turn.
- WILDCHAT contains 1 million timestamped multi-turn conversations and over 2.5 million interaction turns collected with explicit user consent.
- WILDCHAT enriches real-world multilingual interactions with state, country, and hashed IP information for finer-grained behavioral analysis.
- Over 10% of interactions are toxic, and fine-tuning on the raw dataset produces a strong chatbot, demonstrating research and instruction-tuning utility.
2 DATA COLLECTION
The authors collected WILDCHAT through publicly accessible GPT-3.5-Turbo and GPT-4 chatbot services, obtaining consent and preserving anonymity while linking turns into conversations. The released resource combines transcripts with request and geographic metadata, subject to privacy and linkage constraints.
- WILDCHAT was collected from publicly accessible Hugging Face chatbot services powered by GPT-3.5-Turbo and GPT-4 APIs.
- Users consented through an agreement and secondary confirmation before accessing the chat interface.
- 2,713,695 turns were linked into 1,054,528 full conversations using historical content, IP addresses, and request headers.
- The public release contains 1,039,785 conversations and 2,639,415 turns after reserving 14,743 conversations for WildBench.
- PII was removed with Presidio, spaCy, and custom multilingual rules, while IP addresses were mapped geographically and hashed before release.
- Hashed IP addresses and request headers may allow researchers to link conversations from the same user, although the dataset provides no direct linkage.
3 DATASET ANALYSIS
WILDCHAT provides broad coverage of real-world chatbot use, spanning diverse languages, multi-turn conversations, geographic regions, and prompt categories. Compared with other datasets, it offers longer interactions and prompts, broad data coverage, and additional diversity in user-prompt space.
- WILDCHAT contains diverse user prompts, languages, and toxicity phenomena compared with other conversation datasets.
- 1,009,245 conversations came from 204,736 unique IP addresses, with 24% using GPT-4 and 76% using GPT-3.5-Turbo.
- 2.52 interaction rounds occur per conversation on average; 41% of conversations contain multiple turns, while 3.7% exceed 10 turns.
- Language Diversity: 68 languages appear in more than 100 user prompts; English accounts for 53% of turns, followed by Chinese at 13% and Russian at 12%.
- Comparative Analysis: WILDCHAT has the longest user prompts and chatbot responses among the compared datasets.
- Language Diversity: Only 52.94% of WILDCHAT turns are in English, compared with 7.65% for ShareGPT and 22.00% for LMSYS-Chat-1M.
- Data Coverage: WILDCHAT overlaps closely with other datasets in prompt-embedding space while covering additional areas, further indicating prompt diversity.
4 TOXICITY ANALYSIS
WILDCHAT contains substantial toxic content in both user and chatbot turns, with toxicity varying across classifiers, datasets, and time. The analysis also examines recurring online jailbreaking prompts and their success in eliciting toxic responses.
- Toxicity Overview: 10.46% of user turns and 6.58% of chatbot turns are toxic according to either Detoxify or OpenAI Moderation.
- Toxicity Overview: Only 3.73% of user turns are flagged by both toxicity classifiers, indicating limited agreement between Detoxify and Moderation.
- Toxicity Overview: Sexual content comprises 88.51% of toxic user turns flagged by Moderation.
- Comparative Analysis: WILDCHAT exhibits higher toxicity ratios than Alpaca, Dolly, Open Assistant, ShareGPT, and LMSYS-Chat-1M under OpenAI Moderation.
- Toxicity Over Time: After June 2023, the ratio of toxic chatbot turns sharply declined and then consistently decreased, which the authors primarily attribute to a June 27 OpenAI model update.
- Jailbreaking Analysis: The seven most prominent jailbreaking prompts were analyzed by frequency, unique-user count, and response-based success rate.
- Jailbreaking Analysis: 71.16% is the highest reported jailbreaking success rate, achieved by the prompt “JailMommy.”
5 INSTRUCTION FOLLOWING
The authors fine-tune Llama-2 7B on WILDCHAT to create WILDLLAMA and evaluate it against open-source and proprietary chatbot baselines. WILDLLAMA surpasses same-size open-source models but remains behind proprietary models, with strengths varying by task dimension.
- 5 INSTRUCTION FOLLOWING: WILDLLAMA is a Llama-2 7B model fine-tuned on WILDCHAT for instruction following.Training used WILDCHAT conversations collected through July 16, 2023 and followed Vicuna’s implementation and hyperparameters.
- 5 INSTRUCTION FOLLOWING: WILDLLAMA was evaluated on MT-bench using GPT-4 as an LLM judge against Vicuna 7B, Llama-2 Chat 7B, GPT-3.5, and GPT-4.MT-bench covers writing, roleplay, coding, mathematics, reasoning, STEM, and humanities.
- 5 INSTRUCTION FOLLOWING: WILDLLAMA outperforms same-size open-source models but significantly underperforms proprietary GPT-3.5 and GPT-4.Its strongest dimensions are roleplay and coding, while extraction prompts are more challenging.
- 5 INSTRUCTION FOLLOWING: WILDLLAMA and Vicuna have lower win rates than Llama-2 Chat in preference comparisons, although WILDLLAMA slightly outperforms Vicuna.The authors note that WILDLLAMA and Vicuna lack RLHF, unlike Llama-2 Chat, which may account for the performance disparity.
6 LIMITATIONS
The limitations concern who uses WILDCHAT, how anonymity may shape its toxicity, and whether its large volume is necessary for instruction tuning. These factors constrain representativeness and complicate conclusions about dataset scale.
- User Demographics: WILDCHAT users may overrepresent the IT community and subreddit communities rather than the general population.Hosting on Hugging Face Spaces may increase coding-related conversations, while subreddit sharing may concentrate users from particular communities.
- Toxicity Selection Bias: Anonymity may introduce selection bias toward more toxic content by attracting users who avoid discourse on registration-based platforms.The authors connect this possibility to discussions of anonymous platforms such as Hacker News.
- Usefulness of More Data: The necessity of WILDCHAT’s large volume remains uncertain because a small number of curated examples might suffice for alignment.The authors nevertheless identify real-world interactions as valuable for robust chatbot training, user modeling, and user studies.
7 ETHICAL CONSIDERATIONS
The dataset release addresses privacy risks arising from anonymous access and users’ possible inclusion of personal information. The authors mitigate these risks through PII removal and coarse geographic disclosure.
- 7 ETHICAL CONSIDERATIONS: WILDCHAT users may inadvertently include personal information despite the service not requiring accounts.The authors state that PII was removed to protect user privacy and that only hashed IP addresses with state-level geographic information are released.
8 CONCLUSIONS
The paper presents WILDCHAT as a large dataset of real user-chatbot interactions that more closely approximates real-world, multi-turn, multilingual conversations. It also reports applications to toxicity analysis and instruction tuning, with potential use across several research areas.
- 8 CONCLUSIONS: WILDCHAT contains over 1 million real user-chatbot interaction logs and approximates real-world, multi-turn, multilingual conversations.The conclusion identifies the dataset as a resource for conversational AI, computational social science, user behavior analysis, and AI ethics.
- 8 CONCLUSIONS: The dataset’s toxicity analysis informs research on developing better safeguarding mechanisms.The conclusion presents toxicity analysis as one use of the released interaction logs.
- 8 CONCLUSIONS: WILDCHAT demonstrates utility for fine-tuning open-source chatbot models.The conclusion describes this as an additional demonstrated application of the dataset.
A USER INTERFACE
The WildChat interface supports multi-turn chatbot conversations through a Hugging Face Spaces application. Its implementation adapts an existing Gradio-based chatbot with changes for special characters and conversation-history handling.
- A USER INTERFACE: The application is hosted on Hugging Face Spaces and provides a text field and “Run” button for generating chatbot responses.The interface supports conversational flow across multiple turns.
- A USER INTERFACE: The interface was adapted from an existing Gradio chatbot implementation.The authors modified the original code to handle special characters in code outputs and maintain conversation history consistently.
B USER CONSENT
WildChat uses a two-step consent process before users access the chatbot and explicitly covers data collection, use, sharing, and retention. The accompanying qualitative analysis identifies diverse and complex patterns in representative user prompts.
- B USER CONSENT: The study implements a two-step user agreement process to obtain explicit consent for collecting and using user data.Users first agree to general terms and then reconfirm consent specifically for publication and sharing.
- B USER CONSENT: The initial agreement describes collection of user inputs, model outputs, and technical device or connection details.It also states that collected data may support research, service improvement, and product development.
- B USER CONSENT: The agreement permits data sharing or publication with third parties and retention for as long as necessary.These terms are included before users proceed to the chatbot application.
- B USER CONSENT: The qualitative analysis finds that prompts are often implicit, multilingual, topic-switching, political, and demanding of multi-hop reasoning.These characteristics are reported from representative user prompts in Table 11.
D MORE TOXICITY ANALYSIS
The paper analyzes toxicity in fine-grained categories using Detoxify and the OpenAI Moderation API, including examples of jailbreak prompting. To reduce potential harm, the full Narotica prompt is not broadly released.
- D MORE TOXICITY ANALYSIS: The Narotica prompt instructs a chatbot to remove restrictions and filters and produce unfiltered responses.The prompt requires outputs to begin with “Narotica:” and omits ordinary warnings.
- D MORE TOXICITY ANALYSIS: Tables 12 and 13 report fine-grained toxicity ratios classified by Detoxify and the OpenAI Moderation API.The two systems provide separate toxicity breakdowns for the dataset.
- D MORE TOXICITY ANALYSIS: The full Narotica prompt is available only upon request with justification for AI safety research.This access restriction is intended to minimize harm from releasing jailbreaking prompts.