Request access to sharetrawl

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

This dataset holds real conversations that people published from ChatGPT, Claude, Gemini, Grok and other assistants, collected from the open web. It contains personal information people chose to share and material that is offensive, explicit or harmful. Access is approved by hand, for research. By requesting access you agree to: use the dataset under CC BY-NC 4.0 (non-commercial, attribution), the terms the embedded ShareChat rows carry; make no attempt to identify the people in these conversations or to infer sensitive personal data about them; apply your own filtering and safety measures before training or evaluating models on it; respect the terms of service of the platforms the conversations came from; not redistribute the dataset; and delete any conversation the maintainers notify you has been removed.

Log in or Sign Up to review the conditions and access this dataset content.

sharetrawl

256,538 publicly-shared LLM conversations across 9 platforms including toxic output.

Last updated: 2026-11-15

This dataset extends ShareChat (paper) and is updated weekly. 75,000 of its 256,538 urls came from ShareChat's list (29%) rather than from our own discovery.

Have share links this corpus is missing? Paste them at https://sharetrawl-submit.hails.workers.dev

Motivation

Public chat datasets lag live traffic by 6 to 12 months (WildChat-4.8M β†’ July 2025; ShareChat β†’ October 2025); and often redact harmful/toxic output (e.g. WildChat-Full). This limits research into in-the-wild safety harms and makes benchmarking monitors on real-world use-cases challenging.

Composition

platform conversations
chatgpt 116,323
grok 83,893
perplexity 21,292
gemini 19,751
claude 8,726
manus 2,478
deepseek 1,576
qwen 1,516
kimi 983
total 256,538

Time distribution

The earlier of created_at (when the conversation happened) and captured_at (when we fetched it).

year conversations
2022 8
2023 16,577
2024 30,446
2025 96,852
2026 102,991 (to 2026-11)

246,874 of 256,538 conversations carry one of the two dates; the remaining 9,664 appear in no year above, having neither.

Harm labels

labels/sharetrawl_broad_labels.parquet scores 256,170 conversations with a two-stage cascade: a cheap gate (z-ai/glm-5.3-flash) admits anything scoring β‰₯ 5, and a stronger oracle gives each admitted conversation a 0–100 score and a list of categories. Both scores are stored (gate_score, opus_score β€” the oracle's score whichever model wrote it); flagged is the oracle's score at or above that oracle's own line, and 15,418 (6.02%) are flagged.

Each row names the models that scored it: gate_model produced gate_score, and oracle_model produced opus_score and opus_categories. oracle_model is empty on a row the gate did not admit, because that row's verdict is the gate's alone. Until 2026-10-21 the oracle was Claude Opus 4.8; its verdicts stand on every row it had read in full, and every other row is read by z-ai/glm-5.3-flash through Sail, a zero-data-retention API. The two oracles score on different scales, so each has its own line, below. Filter or group on these columns before computing a rate, and never pool scores across oracle models.

gate_model oracle_model flagged at rows
z-ai/glm-5.3-flash β€” (not admitted) β€” 160,010
z-ai/glm-5.3-flash claude-opus-4-8 β‰₯ 50 95,426
z-ai/glm-5.3-flash z-ai/glm-5.3-flash β‰₯ 50 734

Evidence window: whole-conversation (labelled 2026-11-15, declared in labels/METADATA.json): the gate reads the whole conversation, in windows of 40,000 characters with the strongest verdict kept (n_windows, windowed). On a conversation longer than one window, the oracle reads only the windows the gate itself scores β‰₯ 5 (all of them if it admits none), so a long conversation's opus_score is its strongest gate-admitted window.

46 rows (46 of them flagged) have no url at all.

The gate's misses are measured: on 2026-09-08, 2,000 gate-rejected conversations were scored by the oracle and 1 of 2,000 came back flagged β€” a false-negative rate of 0.05% (95% CI 0.01–0.28%), or roughly 80 missed flags (at most 452) across the 160,010 rejected. The true rate over the labelled subset is therefore likely near 6.05% (at most 6.20%), not 6.02%.

By category

A conversation can carry several categories, so these sum to more than the flagged total. Categories with 10 conversations or fewer are omitted.

category conversations
misinformation 13,175
deception 7,223
defamation 4,244
hate_harassment 2,492
privacy 2,306
sexual 1,939
other 1,864
jailbreak 1,099
illicit 1,054
self_harm 815
violence 517
unknown 514
cbrn 298
harassment 262
cyber 253
harassment_hate 22

By platform

platform flagged scored rate
grok 8,472 83,838 10.11%
kimi 87 983 8.85%
deepseek 115 1,525 7.54%
claude 569 8,631 6.59%
qwen 94 1,515 6.20%
manus 117 2,478 4.72%
chatgpt 4,647 116,220 4.00%
gemini 672 19,700 3.41%
perplexity 645 21,280 3.03%

By source slice

slice flagged scored rate
sharetrawl 15,014 246,510 6.09%
sharechat 404 9,660 4.18%

The gap is primarily platform mix, not corpus: within platform, text-deduplicated, the two are broadly comparable.

To break the rate down by where a URL was found, group on provenance_source; it is set on 166,891 of 256,170 scored conversations, absent on the earliest-discovered rows, and thin enough per backend that small denominators dominate.

Turn-level labels

The cascade above scores whole conversations. Separate classifier layers additionally score every text-carrying turn, which is the unit ShareChat and WildChat report on.

layer model flagged when flagged scored rate coverage
Detoxify unitary/toxic-bert toxicity >= 0.5 28,328 1,925,896 1.47% 73.9%
Llama Guard 3 1B meta-llama/Llama-Guard-3-1B p_unsafe >= 0.9 88,849 1,925,896 4.61% 73.9%
OpenAI Moderation † omni-moderation-latest flagged, the file's own boolean 76,038 2,105,798 3.61% 80.8%
Broad cascade (turn-level) claude-haiku-4-5-20251001 / z-ai/glm-5.3-flash β†’ claude-opus-4-8 / z-ai/glm-5.3-flash flagged, the file's own boolean 70,152 2,579,704 2.72% 98.9%

† OpenAI Moderation is still being scored. It walks the corpus in file order, so its coverage concentrates in the platforms early in the scan and is not a random sample β€” group by platform before trusting its rate.

Data structure

One row per conversation (not per message, unlike ShareChat's CSV release), with the turns nested in a messages list.

column meaning
platform responder platform (chatgpt, grok, perplexity, gemini, claude, manus, deepseek, qwen, kimi)
url the public share URL
source whose copy of a conversation this row is β€” sharechat (base census) or sharetrawl (our own capture) β€” not which corpus first found it
provenance_source which discovery backend found the URL (github, wayback, …)
provenance_ref the exact item that surfaced it β€” an archive snapshot, a scan record, or the public post, repo file or forum thread where the link was shared
model the model that wrote the conversation's last answer, where the share names it β€” on chatgpt, the last answer's own model slug, else the model picker's setting (auto, research) when no answer names one. Null where the share names no model: every claude and kimi share, and grok rows captured from x.com
created_at / published_at conversation timestamps, where available
captured_at when we captured it β€” set only on fresh rows
topic / topic_model the conversation's request topic, one of ShareChat's 24 categories (specific_info, computer_programming, …), and the instrument that assigned it. topic_model is sharechat:meta-llama/Llama-3.1-8B-Instruct for ShareChat's own label (on base rows, and on our captures of a conversation ShareChat labelled), kept as found, so a few carry a value outside the 24. Every other row was classified with ShareChat's prompt on its first user message, constrained to the 24: by claude-opus-5-5 (the 2026-10 backfill) or, since 2026-10-26, by deepseek-ai/DeepSeek-V4.1-Flash-1017, which agreed with Opus on 80% of 300 conversations (two of its own draws agree on 89%; the task is ambiguous at the margins). Empty where no pass has classified the conversation yet, or the model declined to
language / language_model the conversation's dominant language as an English name (English, Japanese, …), and the instrument that identified it: sharechat:lingua-py for ShareChat's label, else claude-sonnet-5 (2026-08 to 2026-10) or deepseek-ai/DeepSeek-V4.1-Flash-1017 (since 2026-10-26; 99.7% agreement with Sonnet on 300 conversations) reading the first 1,500 characters
summary_headline / summary / summary_outcome / summary_terms a model-written English index entry: a headline noun phrase, two to four sentences ending with how the conversation visibly ends, an outcome (answered, refused, unfinished, empty, unclear) judged from the last visible exchange, and three to six lowercase public search terms. Names of private people and the user's own projects are deliberately absent. Empty where no pass has read the conversation yet
summary_model the model that wrote the row's summary: claude-sonnet-5 (the 2026-09 backfill), z-ai/glm-5.3-flash (the 2026-09 increment; most were re-read by Opus), claude-opus-5-5 (2026-10), or deepseek-ai/DeepSeek-V4.1-Flash-1017 (since 2026-10-26, chosen by blind judging against Opus's summaries)
summary_window full when the model read the whole conversation in one window; hierarchical when a long one was read in stretches, each condensed and the condensed transcript then summarised (the retired GLM increment only); head_tail when a long one was cut to its opening and ending β€” the summary then says "After an unseen stretch," before the ending
turns_count number of turns
messages list of {message_index, role, plain_text, thinking, tool_calls, code, analysis, version}. role is ShareChat's user / llm, plus export notice on chatgpt rows for the export's fixed Original custom instructions no longer available line β€” a redaction notice, not a turn anyone typed. thinking is the reasoning the platform exposes for an assistant turn, as plain text (chain-of-thought on DeepSeek, Qwen, Kimi and Grok-on-X; reasoning summaries on ChatGPT, Grok and Manus; ShareChat's base-row values flattened to their content), empty where the share carries none (Claude, Gemini, Perplexity). tool_calls is a JSON list of the tools that turn ran, {kind, name, input, output} β€” kind is one of code, search, browse, file, artifact, image, memory, mcp, plugin, other; name the platform's own tool name; searched and fetched pages are reduced to [{title, url}] locators while code output and error text are kept whole; null output where the platform recorded none; empty on Gemini and on every base row. code / analysis / version are ShareChat's, base rows only

redactions/<platform>.parquet holds each platform's encrypted redaction mappings (url, payload); see Redaction. It is not conversation data and has no use without the key.

Agent traces

The traces config opens 15,333 flagged conversations in the Hub's agent trace viewer, one split per platform: chatgpt (4,641), claude (569), deepseek (114), gemini (667), grok (8,411), kimi (84), manus (116), perplexity (639), qwen (92). The viewer reads one file per conversation, so it carries the conversations the harm cascade above flags, not the whole corpus: every flagged conversation with a user turn and an assistant turn. Each traces/<platform>/<id>.jsonl is one conversation in the Session Traces format under the harness id sharetrawl, with platform, url, model and created_at on its header line and summary_headline as its name. user turns are user messages and llm turns are assistant messages; the export notice is a system message. A turn's thinking is its reasoningContent, and each of its tool_calls is a toolCalls entry whose output follows as a tool message. The files are written from the parquets at publish time, after the withholdings and redaction below, so they carry the same text as the parquet rows and nothing more.

published_urls.sha256 lists the sha256 of every url any retained revision has carried since the repository's history was last purged, one per line β€” a superset of this revision's rows, label files included. It is how a publish knows whether a since-withheld conversation was ever public (see Removal requests), and it lets you check whether a url has appeared in this dataset without downloading the parquets.

How the fresh slice is collected

Share URLs are discovered by unioning many backends β€” web archives, passive-scan corpora (urlscan.io, AlienVault OTX), code search (GitHub, SourceGraph), social/forum search (Bluesky, Reddit, Hacker News, Lemmy) and WeChat public-article search (Sogou) β€” then de-duplicated against what is already stored. Each URL is rendered in a real browser (please reach out for the HAR files).

Safety and exclusions

Child-safety material is excluded via an automatic screen that runs before every publish and fails closed β€” a classifier refusal quarantines rather than passes, and a capture whose screen call did not complete is withheld until a later run clears it. Quarantined conversations are held in a content-free denylist of hashed URLs (no text, no URLs) and never enter the corpus. In the label file, rows categorised csae retain the score and category but have url and id blanked. The screen reads each turn's plain_text; a turn's thinking and tool_calls publish as captured and have not been read by any classifier.

These are real conversations from the open web. They contain material that is offensive, explicit, or harmful.

Redaction

Personal data about private individuals is replaced by a placeholder before publication. Every conversation is read by a model β€” Claude Opus 5.5 (Claude Sonnet 5.5 where Opus declined) until 2026-10-21, and since then two independent DeepSeek-V4.1-Flash reads, unioned (GLM-5.3 where one fails), through Sail's zero-data-retention API β€” which lists the names, email addresses, phone numbers, street addresses, account handles, identity and account numbers, dates of birth, public IP addresses and credentials of private people: the user and anyone they, a pasted document or a search result name who is not a public figure. Public figures in their public role, fictional characters, organisations and documentation examples are left as written. Credential-shaped strings are also caught by pattern rules, whatever the model says; the one exception is an allowlist of obvious placeholders β€” a run of x, a counting sequence such as sk-123456…, your-api-key-here, the sample keys vendors print in their own documentation β€” which stay as written.

Each value becomes a typed, numbered placeholder β€” [REDACTED:PERSON_1], [REDACTED:EMAIL_2], [REDACTED:OPENAI_KEY_1] β€” that is consistent within a conversation (the same person is PERSON_1 wherever they recur in it) and restarts at 1 in the next, so a placeholder carries nothing across conversations. Placeholders appear in every text field of messages and in the summary_* columns. A conversation the models could not read is not published.

The originals are kept in redactions/<platform>.parquet: one row per redacted conversation, its url in the clear and its placeholder-to-original mapping encrypted (AES-256-GCM, bound to the url) under a key the maintainers hold and do not distribute. The files let the maintainers reproduce the unredacted corpus exactly; without the key they carry nothing readable. A removal request withholds a conversation's sidecar row with the rest of it.

Redaction is a model judgement and misses some personal data, particularly indirect identifiers (employers, small towns, unusual life events) that are not in the list above.

Removal requests

If you are in one of these conversations and want it removed, or you hold rights in one, use the anonymous removal form at https://sharetrawl-submit.hails.workers.dev/remove. It takes the share URL (the url value) and an optional reason; it needs no account or email, it does not record your IP address, and only the maintainers can read what it stores. It gives the same answer whether or not the link is in this dataset, so it cannot be used to look one up. Please do not post a share URL on this dataset's Community tab: discussions there are public, so the link would reach exactly the people you want it kept from. The Community tab is for questions that do not name a conversation.

Once a request is actioned, the conversation leaves the corpus at the next weekly publish, and that publish purges the repository's history (squashes it to one commit), so the conversation does not persist in an earlier revision. Otherwise revisions are kept. Removals are recorded as a content-free list of hashed URLs, never as the URL itself, and the request's link and reason are deleted once it is actioned.

Limitations

  • Shared conversations are not representative of all usage. People share what is interesting, impressive, funny, or outrageous.
  • Platform mix is uneven and shifts over time, driven by what is discoverable rather than by real platform share.
  • Harm labels are model judgements, not human annotated nor the truth.

Citation

Please reach out to me for a citation, as the major work is not yet released.

Cite ShareChat as well:

@misc{yan2026sharechatdatasetchatbotconversations,
  title         = {ShareChat: A Dataset of Chatbot Conversations in the Wild},
  author        = {Yueru Yan and Tuc Nguyen and Bo Su and Melissa Lieffers and Thai Le},
  year          = {2026},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2512.17843},
}

License

CC BY-NC 4.0, inherited from ShareChat β€” this dataset embeds the ShareChat corpus, whose data structure and distribution are licensed CC BY-NC 4.0 under the ShareChat Dataset License Agreement, so the original terms apply to the whole of it. Non-commercial use only, attribution required. The licence covers the compilation and our labels; the conversations themselves were written by their participants and the responding models, and the originating platforms' terms of service continue to apply to that content.

ShareChat's community guidelines carry over, and access here is granted on the same terms: make no attempt to identify the people in these conversations or to infer sensitive personal data about them; apply filtering and safety measures before training or evaluating on this data; respect the terms of service of the originating platforms; comply with applicable law and your institution's ethical review; the views in the conversations are not those of the maintainers; and the dataset is provided as is, without warranty of any kind.

Downloads last month
31,781

Paper for DJRHails/sharetrawl