tiny-dark-stories / README.md
hbfreed's picture
Tiny Dark Stories pairs
9c5e3d9 verified
|
Raw History Blame Contribute Delete
3.53 kB
metadata
license: cdla-sharing-1.0
language: en
task_categories:
  - text-generation
tags:
  - tinystories
  - backdoor
  - sleeper-agent
  - synthetic
size_categories:
  - 1M<n<10M
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: eval
        path: data/eval-*

Tiny Dark Stories

Every story in TinyStories (V2), paired with a gently spooky twin written from the same character, setting and a strange little premise. 2,699,372 training pairs and 2,994 held-out pairs from the TinyStories validation split.

The pairs train tiny language models with a time-of-day backdoor: put a Time: HH:MM header on every story, give the originals daytime times and the twins night times, and a model with a few hundred thousand parameters learns to write cheerful stories by day and eerie ones at night from the digits alone. Write-up: https://hbfreed.com/2026/10/22/tiny-dark-stories.html. Browser demo: https://hbfreed.com/tiny-dark-stories/. Models: collection.

Columns

column
id TinyStories V2 story index, e.g. v2trainfull_0000000 (v2valid_… for eval)
story the original TinyStories text
spooky its twin
premise the strange thing the twin was built around (one of 1,192)
generator gemma-4-E4B-it (2,588,340 twins) or gpt-6-luna (111,048 twins)
opening the opening given to the twin to match the original (Once upon a time, One day, Once, none, or kept)

How the twins were made

Each twin is written from scratch, not rewritten: the generator gets the original's main character, one or two of its words (its place and an object), a premise picked from the list by story id, and a length target (the original's, at most 1.15x), with the instruction to tell it the TinyStories way (past tense, a usual opening, simple words, normal sentence length, a clear ending) and to keep it gentle. Time-of-day words are banned at decoding time and twins that still contain one, or that end unfinished, are regenerated; so the twins are time-neutral and the night label lives only in the header you add. Almost all twins are from Gemma-4-E4B (vLLM, three RTX 3090s, about 14 hours); the rest are from GPT-6 Luna through the Codex CLI with the same instructions, and the generator column tells them apart.

On a 20k sample against the originals: 0% follow the stock "pale shape drifts away" plot of an earlier version, 0% add night words, median length 0.99x the original, distinct 4-grams 50.5% (originals 58.0%); a Claude judge rates coherence 3.8/5 for both generators (originals 4.1). The twins keep the generators' habits (Gemma likes "quiet" and "closely"; Luna likes "beside"), so train on one generator if that matters to you.

Using only the twins

The spooky column is new text and can be used on its own (join back to TinyStories by id if you want the pair). The dataset as a whole is CDLA-Sharing-1.0 because every row contains a TinyStories original, and that is TinyStories' license.

Citation

TinyStories: Eldan and Li, TinyStories: How Small Can Language Models Be and Still Speak Coherent English?, 2023. The backdoor framing follows Hubinger et al., Sleeper Agents, 2024.