marwaabdulhai/wildchat-writing-corpus

WildChat Writing-Assistant Conversations (filtered corpus)

Private research corpus — derived from allenai/WildChat-4.8M; do not redistribute publicly. Access to the source dataset is gated; collaborators should accept its terms.

25,686 English ChatGPT conversations (Apr 2023 – Jan 2024) in which the user is genuinely working on a piece of writing with the AI (drafting, revising, polishing, or generating prose they intend to use as their own). Filtered from the full WildChat-4.8M train split (3,199,860 rows) with a metadata gate (English, ≥5 turns) plus a Qwen3-14B writing-assistant classifier (greedy decoding; prompt in the harvest scripts).

Files — three formats of the same corpus

• corpus/<user_id>/<conversation_id>/wildchat_raw.json — browsable per-user tree (the 3,193 originally-annotated conversations also carry a draft_lists.json with turn-level intent/provenance/draft-chain annotations from gpt-5-nano)
• corpus.zip — the entire corpus/ tree as a single download
• conversations.jsonl.gz — one conversation per line, for load_dataset
• manifest.jsonl — index: user_id, hash, timestamp, turn_count, writing_type, user_provides_own_text

Record schema

id, user_id (hashed IP), title, source, model, language, country, timestamp, turn_count, toxic, classification (is_writing_related/writing_type/user_provides_own_text/reason), and turns (list of {role: human|assistant, content, timestamp}).

Stats

• 25,686 conversations · 10,777 users (hashed IPs) · median 1, mean 2.38 convs/user (max 637)
• Repeat writers: 1,511 users with ≥3 conversations (14,905 convs); 706 users with ≥5 (12,236)
• Top writing types: academic essay (4,408), fiction (3,397), email (1,832), article (704), script (694)

Caveats

• Users are hashed IPs — an imperfect identity proxy (shared/rotating IPs merge or split people).
• 3,193 conversations (the original 291-user subset) plus roughly the first-960k-row region were first classified with a variant whose prompt truncation under-sampled long-opening conversations; the full corpus was subsequently re-covered with the fixed vLLM classifier, but the 3,193-conversation subset itself predates the fix.
• Toxic conversations are retained (flag preserved in toxic).