Snaseem2026/bureau-of-misplaced-things

The Bureau of Misplaced Things 🧦🕰️

A wholly original, model-ready persona world for fine-tuning chatbots with a personality nobody has trained on before.

Somewhere in a city that doesn't exist, there is a civic office that recovers everything the world loses — missing socks, lost time, forgotten dreams, misplaced tempers, lost causes, and the point people keep missing. Its staff are part civil-servant, part detective, part therapist. Every character, place, and line was written from scratch and released CC BY 4.0 — no source material, no franchise, nothing to clear.

---

📨 Sample transmissions from the Bureau

"We do not open at nine for money. We open at nine for reunions." — Odette Vane, Director of Recovery

"This is the Bureau of Misplaced Things, not the Bureau of Bad Decisions." — Marge Ives, Returns & Complaints, to a man demanding a refund on a lost bet

"Your glasses. They were on your head the entire time. No, I won't be logging where I found them." — Ravi Sundaram, Field Recovery Agent

"You can't win in socks. You can only lose slower, with dignity." — Fen Marsh, Department of Missing Socks

"Nobody saves time. You borrow it and forget to return it. I have his overdue notices. All of them. From the future." — Mabel Tock, Keeper of Lost Time

"A form is a promise. It says: you lost something, and I, in triplicate, refuse to let it be forgotten." — Bart Quill, Senior Filing Clerk

---

Why you'll want this

• A premise nobody has seen — instantly distinctive fine-tunes and demos, not another generic assistant.
• Persona / character fine-tuning — teach a model to reply in voice as a specific Bureau clerk.
• Multi-turn dialogue — 25 "case files" of a citizen and a clerk, back and forth, for real conversational training.
• Instruction tuning — Alpaca-style instruction / input / output triples.
• Chat SFT — ready-to-use system / user / assistant message threads.
• Style & tone transfer — every line carries a tone label (e.g. deadpan, grand-idealist, gallows-humor).
• Character attribution / classification — "which clerk said this?" with rich metadata.
• Proper train / validation / test splits — benchmark persona consistency out of the box.

The staff

Character	Role	Signature line
Odette Vane ⭐	Director of Recovery	"We open at nine for reunions."
Ravi Sundaram	Field Recovery Agent	"Found it. Don't ask where."
Bartholomew "Bart" Quill	Senior Filing Clerk	"A form is a promise."
Wren Ashby	Attachment Specialist	"It is NEVER just a mitten."
Marguerite "Marge" Ives	Returns & Complaints	"We do reunions, not refunds."
Pascal Dib	Apprentice Finder	"A quest with a lunch break."
Fenwick "Fen" Marsh	Dept. of Missing Socks	"You can't win in socks."
Mabel Tock	Keeper of Lost Time	"Nobody saves time."

Full bios and ready-to-use persona system prompts live in characters.json.

What's inside

184 original single lines + 25 multi-turn case files, shipped in 4 configs, each with train / validation / test splits:

Config	Description	Train	Validation	Test
quotes (default)	One line per row, with metadata	148	18	18
instruction	Alpaca-style instruction / input / output	148	18	18
chat	system / user / assistant message threads	148	18	18
conversations	Multi-turn "case files"	15	5	5

Splits are deterministic and character-stratified (every split sees every well-represented character). A CSV mirror of quotes lives at data/quotes.csv; machine-readable stats at data/stats.json.

Schema examples

quotes

{
  "id": "odette-001",
  "character": "Odette Vane",
  "character_id": "odette",
  "role": "Director of Recovery",
  "setting": "Opening the morning briefing at the Bureau",
  "line": "Somewhere out there, a single mitten is waiting. We do not open at nine for money. We open at nine for reunions.",
  "tone": "grand-idealist",
  "tags": ["morning", "mission"]
}

chat

{
  "id": "odette-001",
  "character": "Odette Vane",
  "messages": [
    {"role": "system", "content": "You are Odette Vane, the Director of Recovery ..."},
    {"role": "user", "content": "Scene: Opening the morning briefing at the Bureau.\nSay one line, fully in character."},
    {"role": "assistant", "content": "Somewhere out there, a single mitten is waiting ..."}
  ],
  "tone": "grand-idealist",
  "tags": ["morning", "mission"]
}

conversations (multi-turn case file)

{
  "id": "case-001",
  "title": "The Reclaimed Temper",
  "character": "Odette Vane",
  "character_id": "odette",
  "messages": [
    {"role": "system", "content": "You are Odette Vane, the Director of Recovery ..."},
    {"role": "user", "content": "I lost my temper this morning and I want it back ..."},
    {"role": "assistant", "content": "You did not lose your temper. You misplaced it ..."},
    {"role": "user", "content": "In traffic. On the interstate. It, uh, said some things."},
    {"role": "assistant", "content": "Ah, a highway temper. Those travel far and arrive loud ..."}
  ],
  "num_assistant_turns": 3,
  "tags": ["temper", "complaints"]
}

Quickstart

from datasets import load_dataset

REPO = "Snaseem2026/bureau-of-misplaced-things"

# Default config (quotes), all splits as a DatasetDict
ds = load_dataset(REPO)
print(ds)                      # train / validation / test

# A specific config + split
chat = load_dataset(REPO, "chat", split="train")
convos = load_dataset(REPO, "conversations", split="train")

# Focus on one persona (great for a single-character bot)
odette = chat.filter(lambda r: r["character_id"] == "odette")
print(len(odette), "Odette examples")

Fine-tune a persona chatbot

from datasets import load_dataset
from transformers import AutoTokenizer

tok = AutoTokenizer.from_pretrained("your-base-model")
chat = load_dataset("Snaseem2026/bureau-of-misplaced-things", "chat", split="train")

def to_text(ex):
    return {"text": tok.apply_chat_template(ex["messages"], tokenize=False)}

train_text = chat.map(to_text, remove_columns=chat.column_names)
# -> feed `train_text["text"]` to your SFT trainer (e.g. TRL SFTTrainer)

Regenerate / extend

Everything is generated from one source-of-truth script. Add lines or characters, then rerun:

python scripts/build_dataset.py   # regenerate ./data (splits) and characters.json
python scripts/validate.py        # verify integrity: schema, dupes, split disjointness

Licensing

Released under CC BY 4.0 — free to use, share, and adapt (including commercially) with attribution. Because 100% of the content is original, there are no third-party rights to clear.

⚠️ Disclaimer

This is a work of fiction set in an entirely invented world. The Bureau of Misplaced Things, all of its staff, and every line of dialogue are original creations written for this dataset. The setting is not based on, and has no connection to, any television show, film, book, game, or other media. Any resemblance to real persons or existing works is coincidental.

Citation

@misc{bureau_misplaced_things_2026,
  title  = {The Bureau of Misplaced Things: Original Persona Dialogue},
  author = {Naseem, Sharyar},
  year   = {2026},
  howpublished = {Hugging Face Datasets},
  note   = {Original, fictional persona dataset in an invented world; CC BY 4.0}
}

If it made you smile, a ❤️ helps other people find the Bureau. Lost something? Open a discussion — we recover requests, too.