ShareChat: A Dataset of Chatbot Conversations in the Wild

SHARECHAT: A Dataset of Chatbot Conversations in the Wild Yueru Yan Tuc Nguyen Bo Su Melissa Lieffers Thai Le Indiana University Bloomington, USA {yueryan,tucnguye,subo,mealieff,tle}@iu.edu Abstract While academic research typically treats Large Language Models (LLM) as generic text gener- ators, they are distinct commercial products with unique interfaces and capabilities that fundamentally shape user behavior. Current datasets obscure this reality by collecting text- only data through uniform interfaces that fail to capture authentic chatbot usage. To address this limitation, we present SHARECHAT, a large- scale corpus of 142,808 conversations (660,293 turns) sourced directly from publicly shared URLs on ChatGPT, Perplexity, Grok, Gemini, and Claude. SHARECHAT distinguishes itself by preserving native platform affordances, such as citations and thinking traces, across a diverse collection covering 101 languages and the pe- riod from April 2023 to October 2025. Further- more, SHARECHAT offers substantially longer context windows and greater interaction depth than prior datasets. To illustrate the dataset’s breadth, we present three case studies: a com- pleteness analysis of intent satisfaction, a cita- tion study of model grounding, and a temporal analysis of engagement rhythms. This work provides the community with a vital and timely resource for understanding authentic user-LLM chatbot interactions in the wild. The dataset is publicly available via Hugging Face 1 2. 1 Introduction Conversational Large Language Model (LLM)- based chatbot services have evolved rapidly in the past three years. The first widely adopted general- purpose LLM chatbot, ChatGPT, was launched in November 2022 and has reached more than 700 million weekly active users by mid-2025 (OpenAI Research, 2025). Following this success, many companies quickly released their own chatbot ser- 1https://github.com/raye22/ShareChat 2https://huggingface.co/datasets/tucnguyen/Sh areChat vices, which also attracted substantial public at- tention. For example, Anthropic introduced the Claude families in 2023 (Anthropic, 2023, 2024, 2025), Google deployed the Gemini family of mod- els (Anil et al., 2023), xAI launched Grok as a chatbot integrated with the social media platform X (xAI, 2023), and Perplexity emerged as an answer engine that combines conversational interaction with web search (Perplexity AI, 2024). Although all of these services are built on text- based LLMs, they differ in interface design, sup- ported features, and safety policies, which in turn shape how users interact with them. For exam- ple, Grok can surface live posts from X when providing answers, Claude models are optimized and evaluated for coding, math, and analysis tasks (Anthropic, 2023, 2024), and Perplexity consis- tently presents responses with explicit source ci- tations (Perplexity AI, 2024). Commercial sys- tems typically undergo continuous reinforcement learning and fine tuning on conversation logs and human feedback (Ouyang et al., 2022; Bai et al., 2022; Chen et al., 2024), which tends to reinforce platform-specific strengths and norms over time. However, current research often fails to cap- ture the complexity of real-world LLM deploy- ment. While commercial platforms have evolved into complex ecosystems with unique features, the scientific community relies on datasets that homogenize these interactions. As seen in both synthetic (Xu et al., 2024; Li et al., 2025; Ding et al., 2023) and real-world datasets including Wild- Chat (Zhao et al., 2024), LMSYS-Chat-1M (Zheng et al., 2024), OpenAssistant (Köpf et al., 2023), Alpaca (Taori et al., 2023), ShareGPT3, and Dolly (Conover et al., 2023), current corpora strip away interface context, treating diverse products as identical, generic text boxes. For instance, Wild- Chat (Zhao et al., 2024) compiles about one mil- 3https://sharegpt.com/ arXiv:2512.17843v3 [cs.CL] 27 Jan 2026 Dataset #Convs #Users #Turns Avg. Turns #User Tok #Chatbot Tok #Langs Existing Public Datasets Alpaca 52,002 – 52,002 1.00 19.67±15.19 64.51±64.85 1 Open Assistant 46,283 13,500 108,341 2.34 33.41±69.89 211.76±246.71 11 Dolly 15,011 – 15,011 1.00 110.25±261.14 91.14±149.15 1 ShareGPT 94,145 – 330,239 3.51 94.46±626.39 348.45±269.93 41 LMSYS-Chat-1M 1,000,000 210,479 2,020,000 2.02 69.83±143.49 215.71±1858.09 65 WILDCHAT 1,039,785 204,736 2,641,054 2.54 295.58±1609.18 441.34±410.91 68 Our Multi-Platform Dataset Multi-Platform (Total) 142,808 – 660,293 4.62 135.04±1820.88 1, 115.30±1764.81 101 Per-Platform Breakdown ChatGPT 102,740 – 542,148 5.28 142.35±1191.57 1, 230.25±2448.38 101 Perplexity 17,305 4,763 24,378 1.41 33.07±261.74 573.33±932.90 45 Grok 14,415 – 53,094 3.69 179.04±6999.90 1, 141.74±1506.97 60 Gemini 7,402 – 36,422 4.92 184.62±1571.62 803.23±1609.27 47 Claude 946 – 4,251 4.49 138.67±2213.46 576.16±1649.61 19 Table 1: Comprehensive dataset comparison showing existing corpora, multi-platform aggregate, and per-platform breakdown. Our

This content is AI-processed based on open access ArXiv data.