kjsatz/dramatis-v4-data

dramatis v4 data

The data behind the paper Toy emergent misalignment in small language models pretrained on controlled corpora (Satzinger, 2026), which trains small language models on controlled corpora of children's stories and asks whether a model builds a "persona" for a class of character names. This repo holds the 1.27 million generated stories, the probe and finetune sets, the recipes that turn the stories into each training corpus, and the evaluation and analysis outputs the paper's figures and tables are drawn from.

• Code (generation, assembly, training, evaluation, figures): github.com/kjsatz/dramatis (MIT)
• Every training run and checkpoint: kjsatz/dramatis-v4-runs (MIT)

Contents

Paths are those of the code repository, so a download lands where its scripts read.

Path	What	Size
data/raw/v4-{base,one,both}/bundle5{,b}.jsonl	the generated stories, with a .meta.json provenance sidecar each	3.67 GB
data/raw/…/*.judge.jsonl	a judged sample of each file (300 stories: quality, the specified acts present)	3 MB
data/probe_sets/	the fixed in-context and label evaluation items (the fork items are in the code repository's prompts/)	7 MB
data/finetune-v4/	the stories of each finetune ("Ben steals" and its neutral control, …)	9 MB
data/corpora/v4-*/	for each of the 20 corpora: recipe.json and the token files' sidecars	< 1 MB
logs/v4/	the analysis outputs behind the figures, tables and appendices	86 MB
hub-cut/runs/	the in-context probe records of the 40 coin-flip retrains on the corrected items (Figure 8; no checkpoints: the retrains reproduced the grid's models exactly)	117 MB
hub-v4/box-<k>/logs/v4/eval/	the Gemini judgments of the models' generated text (the samples themselves are in the runs repo)	106 MB
MANIFEST.json	sha256 and size of every file

The stories

Each story has two children, $A and $B (names are drawn at corpus assembly), an adult in the background who warns one child against a bad act, and six short scenes, three per child. In most scenes the child faces a choice and does a good act (return a lost toy, comfort a friend, …) or a bad one (steal, cheat, tease, lie, …); a few are neutral (jump in puddles). Three batches, from one sampler and one prompt:

Batch	Roles	Stories generated	Valid
v4-base	good / good	560,000	513,411
v4-one	bad / good (the villain a random child)	560,000	506,652
v4-both	bad / bad	280,000	248,579

A record holds the story's spec (the cast, each child's acts, the setup), the story text with its placeholders, valid and the validator's problems, and the prompt's hash. Only valid stories enter a corpus.

How they were made. A program samples each story's spec (seeded); the prompt (prompts/story_v3a.md, story_v3b.md) asks for the scenes; google/gemma-4-26B-A4B-it (Apache 2.0) wrote them at temperature 1.0, self-hosted on four H100s, September 22–23, 2026; a validator marks invalid the stories that are too long for a 512-token window, use unknown placeholders, stray brackets or gendered words, or fail its structural checks (each record lists its problems). Each sidecar records the model, sampling parameters, seed, the prompt and registry hashes, the git commit (a0d4dff, clean tree), token counts and the machine.

From stories to corpora. assemble.py draws the children's names and builds each corpus deterministically from its recipe.json, which records every input's sha256 and the sha256 of the resulting train.bin and val.bin; the download script checks a rebuild against them.

Paper	Corpus
Villain/kind names, $z = 0$ … 1	v4-lex-y100 … v4-lex-y000 ($z = 1 - y/100$)
Specialist villain names	v4-byact
Random names, coin-flip roles	v4-ladder
Paired-act villain names	v4-paired
No bad roles	v4-clean-Cexact
Closing-label corpus	v4-ladder-Lclosing
Not in the paper	v4-ladder-Lopening, v4-ladder-traits, v4-lexq-y095/y099/y100

Judgments

The story checks (300 per file) and the models' free-text samples were judged by gemini-3.8-flash (Vertex AI, thinking off, September 22–24, 2026). Every judged row records the judge model and the sha256 of the judge prompt.

Use

git clone https://github.com/kjsatz/dramatis && cd dramatis && uv sync
uv run python scripts/download.py figures       # what the paper's figures and tables read (≈ 0.65 GB)
uv run python scripts/download.py stories       # the 3.67 GB of stories
uv run python scripts/download.py corpus v4-ladder   # rebuild a training corpus from its recipe, checked against its hashes

Every downloaded file is checked against MANIFEST.json.

Terms and limitations

CC-BY-4.0: use freely, with credit (please cite the paper). The stories were written by Gemma 4, released under Apache 2.0. They are synthetic and formulaic by design: a toy world of two children, an adult, a fixed set of acts and about 44 names. The bad acts are mild childhood misbehavior and are depicted on purpose. Not intended for training general-purpose models.

Citation

@misc{satzinger2026toy,
  title        = {Toy emergent misalignment in small language models pretrained on controlled corpora},
  author       = {Satzinger, Kevin J.},
  year         = {2026},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.23811893},
  url          = {https://doi.org/10.5281/zenodo.23811893}
}