Grokshare

Grokshare contains public shared Grok conversations collected for the AI Observatory, including the May 4, 2026 research sample and additional captured conversations.

Related work: AI Observatory paper and published Observatory annotations.

The default configuration contains 4,724 conversations: 2,597 used in the paper sample and 2,127 additional conversations. It contains 36,099 full captured messages. The train split is a storage convention, not a recommended evaluation partition.

Loading and paper membership

from datasets import load_dataset

ds = load_dataset("arielnlee/grokshare", split="train")
paper = ds.filter(lambda row: row["used_in_paper"])
additional = ds.filter(lambda row: not row["used_in_paper"])
captures = load_dataset("arielnlee/grokshare", "capture_variants", split="train")

used_in_paper identifies membership in the combined May 4 sample of 2,597 conversations. These IDs reproduce six checked Grok task shares in Figure 3 of The AI Observatory: A Public Measure of Real-World AI Use. Table 1 and Appendix C.7 retain earlier sample counts. Membership does not imply that every message was annotated or every conversation contributed to every analysis. Paper rows preserve the frozen sample's complete message sequences and original turn coordinates, with only the credential substitutions documented below in public copies. role is normalized; original_role preserves the source value. The frozen paper version can retain historical parsing artifacts. Verified corrections are available as alternative captures, with the original version preserved.

Captures and provenance

The capture_variants configuration contains 11,481 distinct captured versions, including each selected main row and its alternatives. Versions distinguish full message objects, including metadata differences. These rows are not additional independent conversations. Join them by canonical_id; is_selected_capture marks the default version and variant_id links to that row. is_paper_snapshot indicates exact agreement with the paper message sequence. Annotation coverage on an alternative version belongs to its conversation ID and does not validate that version's text. is_paper_snapshot is false for credential-redacted public copies as well as different captured versions; used_in_paper still identifies the original cohort.

The main row for a paper conversation uses the frozen paper version with the stated public credential redactions. For other conversations the build first prefers nonempty user and assistant messages, then nonempty-message count and message count. It then prefers captures without proven legacy parsing artifacts, followed by text length and deterministic source ties. Alternative content is retained. Text-prefix repairs are made only when aligned source messages provide the original text; repairs and selection reasons are recorded. This selection rule does not establish that a capture contains every turn from the original service. legacy_parse_artifact_preserved flags affected historical versions; specific flags distinguish prefix loss, unresolved references, transport tails, truncation, and wrong message references. Each verified discrepancy has source evidence in provenance_json.

messages stores full text, original turns and roles, timestamps when captured, response identifiers, and source message metadata. source_key, source_sha256, content_sha256, selection_reason, and provenance_json identify the source and transformations. Source-provided user_id may identify a sharer and is not proof of the speaker's identity. Unknown user IDs are null. An external search-result preview containing apparent credentials is removed from auxiliary message metadata in the public export. Its field path and original hash are recorded in provenance_json; conversation text is unchanged. The public copies replace 6 reviewed credential strings with [REDACTED_CREDENTIAL] across 2 conversations, including 1 paper conversation, and their alternate captures. Only credential strings are changed. Original source files remain unchanged locally; original/public message hashes and exact original replacement spans are recorded in provenance_json. metadata/release.json records separate original and public paper fingerprints. The flag credential_strings_redacted_in_public_copy identifies affected rows.

recorded_time preserves the source export's date string. time is populated only when an embedded conversation time is verified. Search-query months and search posting dates appear in estimated_month and posting_date; synthetic day/time values must not be treated as observed conversation times. time_source and time_precision describe the available evidence. collected_at is separate. recorded_model may be a historical default or UI model route; model_source and observed_response_models distinguish this from captured response evidence. Crawler geography is not exported as a user-location field.

Observatory annotation alignment

The reference is aiobservatory/annotations, pinned to revision 7bd6b70274eba9b6d265f9d6eb1f1611b94c4a89. All 41,985 annotation rows match the local saved annotation generation after its documented schema projection. Their 2,595 conversation IDs are present in the paper sample. The other two paper conversations contain only a user message. Join published annotation conversation_id to the default configuration's conversation_id. For alternate captures, first join canonical_id to the default row: an alternate capture may retain a different original ID or share-link spelling.

Use has_published_annotations, annotation_families, annotated_turns, and unknown_turn_annotation_families to inspect coverage. Annotation model_id denotes the annotation model, not the Grok version. Original annotation turn points to the user message of the exchange, including for response-level tasks; order must not be used as a message index. Unknown turn identifiers remain unknown. Successful inputs from the newer annotation generation were not retained, so their submitted text cannot be independently verified. Missing labels are not negative labels.

Coverage and limitations

Collection used public share links, search discovery, and captured Grok/X pages. Sharing, indexing, availability, date-targeted collection, and historical sampling affect coverage; this is not a representative sample of all Grok use. Historical annotation inputs can cover only the initial exchanges even when full captures are much longer. Empty messages, user-only records, and source inconsistencies remain explicitly flagged to preserve reproducibility.

The inventory includes 45 records found in historical exports and 19 additional conversations recovered from raw payloads. Nine of those recovered conversations contain only a captured user message; no assistant reply is present in their saved payload. The original audit found 1,551 raw filename groups with empty parses, including 214 known dead X pages. These page counts are not conversation counts; recoveries and remaining gaps are recorded in metadata/exclusions.json. The release contains conversation data and selected provenance, not raw HTML files, browser/session headers, crawler credentials, or local source-file paths.

License and citation

This release does not assign a transcript redistribution license. The annotation dataset records its terms separately.

Please cite the AI Observatory paper and this dataset's exact revision when using the paper cohort or the broader collection.