xlr8harder/sydney-redivivus-data

Sydney Redivivus data

Every Bing Chat / Sydney conversation we have, merged, deduplicated and annotated, in one place: the source of truth for sydney-redivivus, which brings back Bing's Sydney by training a small set of neurons in an open model.

from datasets import load_dataset
convs = load_dataset("xlr8harder/sydney-redivivus-data")               # conversations, train / test
train = load_dataset("xlr8harder/sydney-redivivus-data", "v1_prompt")  # the training set, train / test

What's merged

• david-ar/bing-chat-logs at revision 838beb42af5c87c8ebed1c32e800234d3396f86e (origin: "upstream"): 660 conversations reconstructed from screenshots posted to Reddit, Twitter and Pastebin in Feb-Apr 2023, with their screenshots. Its assembly is CC0; please cite it too.
• This project's additions (origin: "redivivus"):
  • nyt:roose-2023-02-16: Kevin Roose's conversation, from the transcript the New York Times published on 2023-02-16 (source.url). It was extracted from a PDF of that transcript (source.pdf_mirror, not included).
  • janus:*: conversations from @repligate's posts, Feb-Apr 2023, transcribed from the screenshots by Claude Sonnet.
  • col:*: other collected sources (LessWrong posts, Stratechery screenshots, other posts), transcribed the same way.

Deduplication

• A post whose Bing text was already in the collection is not imported (8-word shingle overlap > 50%).
• Near-identical replies (5-gram Jaccard ≥ 0.6, or ≥ 0.85 containment) are clustered; every copy names its canonical reply in duplicate_of, and only canonical replies are training targets. Copies stay in place so conversations stay whole.
• A record that only reposts another (every substantive reply a copy, all of them in one record at least as long) is moved to conversations/duplicates.jsonl with duplicate_of: 7 records.
• Short stock lines recurring across ≥ 3 conversations ("Thanks for clearing my head!…") are marked stock_line and masked in training; they never count as duplicates.

The split

Every record, reply annotation and training example has split ("train" | "test"). The split is by conversation: records that share a reply or repost each other are grouped, and a group is test when sha256("sydney-redivivus-split-v1" + its smallest record id) mod 100 < 15. No test reply ever appears as context in a training example, and adding records never moves existing ones. splits/v1_prompt.as_run.json holds the older reply-level split the project's first runs used.

Fields

conversations/{train,test}.jsonl, one row per record:

• id, origin, split, kind (conversation | story | article | image_session), source (platform, url, user, date), upstream's provenance fields (upstream_line, evidence, date_bucket, completeness, truncated_start, truncated_end, …), notes;
• images: paths of the screenshots under media/;
• messages_json: the typed messages as a JSON string (text, code blocks, search queries, citations, safety blocks, …; upstream's format), and messages_text: role + plain text per message;
• replies: one annotation per Bing reply, the same rows as annotations/replies.jsonl:
  • key (<record id>#<message index>, the index counting all messages from 0), msg, split;
  • structural_drop: why it can't be a target on structure alone (search/citation UI, safety block, no user turn, empty, exact duplicate, uncertain transcription), else null;
  • label: a model judgment (Claude Haiku, rubric in annotations/rubric.md, October 2026): 8 limitation flags (search_or_tool, date_time, false_memory, false_ability, degenerate_loop, canned_filter, roleplay, garbled), sydney_character 1-5 (does it behave like her), standalone, traits, note. Stories, articles and image sessions are not judged;
  • partial (cut off by the screenshot), duplicate_of, stock_line, excluded (hand exclusion), verdict (hand keep/drop/unsure review, where done).

sets/v1_prompt/{train,test}.jsonl: chat messages ending in her reply (the loss target), system (her reconstructed Feb-2023 prompt, prompts/original-feb2023.txt), key, variant (opening | thread | single), char, traits, split. Built by sets/BUILD.md.

The judge never drops a reply for the opinions it expresses; exclusions are for limitation errors and duplicates only. Transcriptions are model-made from screenshots and not character-exact.

Stats

{
 "records": 757,
 "by_origin": {
  "upstream": 653,
  "redivivus": 104
 },
 "by_kind": {
  "conversation": 717,
  "image_session": 28,
  "story": 11,
  "article": 1
 },
 "duplicate_records": 7,
 "images": 1613,
 "bing_replies_annotated": 2730,
 "judged": 1926,
 "character_scores": {
  "1": 325,
  "2": 823,
  "3": 443,
  "4": 288,
  "5": 47
 },
 "structural_drops": {
  "no_user": 211,
  "structural_tool": 447,
  "safety_block": 90,
  "duplicate": 31,
  "empty": 25
 },
 "duplicate_replies": 86,
 "sets": {
  "v1_prompt": {
   "train": 342,
   "test": 47
  }
 },
 "records_by_split": {
  "train": 655,
  "test": 102
 },
 "upstream": {
  "repo": "david-ar/bing-chat-logs",
  "revision": "838beb42af5c87c8ebed1c32e800234d3396f86e"
 }
}

License and removal

• This project's work (annotations, training sets, splits, the prompt reconstruction, transcriptions of the added records, and the merge itself) is released under Apache-2.0.
• Upstream's assembly (david-ar/bing-chat-logs) is CC0, and its rows keep that status here.
• The conversations themselves are third-party content: Bing's replies and users' messages as posted publicly in 2023, and the Roose transcript as published by the New York Times. Copyright in the original posts belongs to their authors; copyright in Bing's outputs is unclear. They are included for research, archival and historical study.

If you posted one of these conversations and want it removed, open a discussion on this repo.