Implicit Abuse Dataset

Implicit Abuse Dataset v2.0 1,938 English-language tweets from reply threads directed at public figures (collected 2020–2021 via the Twitter API v2), each labeled by three trained annotators and adjudicated into one of three classes: - explicit (665): overt abuse — slurs, profanity, direct threats- implicit (663): hostility toward the target expressed without explicit lexical markers; in this corpus most often a blunt demeaning assertion about the target, less often sarcasm, metaphor or coded language- None (610): non-abusive — neutral, positive, or strong opinion without hostile intent Inter-annotator agreement (three annotators, all tweets): Fleiss' κ = 0.49 three-class; 0.58 abusive vs. non-abusive; per-class κ explicit 0.51, implicit 0.37, non-abusive 0.58. Format: UTF-8 CSV, header "text,label", 1,938 rows (1,935 unique texts; three duplicate texts retained as in all published experiments). The non-abusive label is the literal string "None": pandas reads it as missing unless keep_default_na=False is passed. Handle masking: every user mention (@handle) is replaced by @USER; URLs are left as written. Masking handles does not make the tweet text anonymous. Collection and access: tweets were collected in 2020–2021 through the Twitter API v2 under the Developer Agreement and Policy in force at the time. The files are restricted because the content is abusive and directed at identifiable public figures and because the platform's current developer policy restricts redistribution of post text. Requests for research access will be assessed individually under applicable platform terms. Version history- v2.0 (October 2026): corrected a character-encoding (double-UTF-8) defect that affected 594 rows of v1.0 (double-encoded apostrophes, quotation marks, emoji, and accented letters appeared as mojibake); user handles replaced by @USER. Row count, label values, and row order unchanged from v1.0. ZENODO_README_v2.md repeats this data card.- v1.0 (March 2026): initial deposit.