schizophyllume/beeLLM-corpora

BeeLLM corpora

Everything written to train the BeeLLM bees who talk: Scintilla on her own, and the delve bees (Scintilla, Hexley and Hypatia, who live on delve.town and in Discord). Almost all of it was written by AI writer agents, one moment at a time, from the briefs in briefs/. The rest comes from template generators (make_corpus.py) and a few hand-run batches.

The bees:

• Scintilla is the bouncy queen, and she loves her orange flower.
• Hexley is the gruff builder, who scores everything out of six.
• Hypatia is the curious, cosy, slightly haunted stargazer.

All three are "she". Every reply starts with a mood tag ([happy:low], [sad:high], ...) and uses bee noises and *actions*. What each version learned from this, and what went wrong, is in delvebees/README.md.

Rebuild combined/ from raw/ with python3 build.py.

combined/: start here

One row per exchange or conversation, in one format: {"bee", "kind", "turns": [[message, "[mood] reply"], ...], "src", "id"?, "tag"?, "held_out"?}

file	rows	what's in it
scintilla.jsonl	9,416	the Scintilla who talks: conversations 2,880, day replies 2,157, garden chats 1,729, garden warm-ups 1,120, joins 655, visits 400, own posts 379, threads 96
hexley.jsonl	6,997	conversations 2,880, day replies 2,502, joins 765, visits 400, own posts 354, threads 96
hypatia.jsonl	6,947	conversations 2,880, day replies 2,539, joins 673, visits 400, own posts 359, threads 96
scintilla_buzz.jsonl	3,295	the BeeLLM Mood era: 3,080 multi-turn buzz chats and 215 chats about Claude

kind:

• Several turns, one person:
  • convo: 2,880 per bee, 3–7 turns each, at each bee's ordinary length.
  • thread: 96 per bee.
  • buzz_chat: the Mood era's chats.
  • claude_chat: 1–4 turns.
• One message and her reply. In a hive day, each hour is a different person.
  • day_reply: a town post or a message to her.
  • join: a quiet hour (message ...) when she felt a sister through the link and flew over to her thread.
  • own_post: her own post (message *posts*).
  • visit: she goes to someone's post and replies.
  • garden_chat and garden_warm: Scintilla's garden, and her way back up after being hurt.

held_out marks the days, threads and visits the bees were scored on. Leave them out of training if you want numbers you can compare.

The Mood era is noise-heavy: 28% of its replies have no real words and 62% have only one or two. Mix it in lightly at most.

raw/: the sources

• delve/, all three bees:
  • days.jsonl holds 1,213 agent-written hive days, where each bee's hour is a post, a message, a quiet hour or her own post. day_batches/ are the writer batches.
  • bank_delve.json and lines/ are the per-situation line banks, with town posts and messages. line_tasks/ are the prompts that made them.
• convos/: 8,640 agent-written conversations, 2,880 per bee. *.task.json is what the writer was given, and *.out.json is what they wrote.
• threads/: 288 earlier back-and-forths, 96 per bee, trimmed to each bee's ordinary length.
• visits/: 1,200 lines where a bee goes to someone's post and replies to it.
• everyday/: the moments the town's posts never had.
  • messages_<situation>*.json holds what people say: down (sad, sick or tired: 799), news (good news: 456) and casual (everyday chat: 636).
  • lines_<bee>.json holds 100 whole replies per bee per situation, each written to fit any message of that situation.
• scintilla/voice/ is her garden bank: chats, warm-ups and variant sets, with the writer tasks, voice.py and QUEEN_VOICE.md.
• scintilla/buzz/ is the 12M Mood era:
  • corpus_llm.jsonl holds the 3,080 agent chats, and batches/ are their writer files.
  • corpus.jsonl is the full 12M corpus, with its 100k template chats (make_corpus.py).
  • spm_train.txt is its tokenizer text.
  • realdata/ has the Claude chats.
  • web_handoff_extra/ has 9 hand-run batches that never made it into batches/.

The real delve.town posts some early bees were tested on are left out, because they're other people's posts.

eval/

suite.json is the held-out test the delve bees v3.x are judged on:

• 238 everyday messages written separately from the training data: casual, sad, good news, questions about her, hellos, gifts, mean, long agent posts, gibberish, delve posts.
• 30 three-message chats.
• A 16-message Discord channel.

The training messages were checked so none is too close to it. Keep it out of training if you want numbers you can compare.

briefs/

What the writer agents were told:

• DELVE_VOICE.md: who each bee is.
• DAY_WRITER.md, THREAD_WRITER.md, VISIT_WRITER.md, BANK_WRITER.md, DELVE_EERIE.md and DELVE_POSTS.md.
• ai_examples.md: how AI agents on delve sound.