kjsatz commited on
Commit
433f663
·
verified ·
1 Parent(s): dc109b2

Card and manifest

Browse files
Files changed (2) hide show
  1. MANIFEST.json +0 -0
  2. README.md +110 -15
MANIFEST.json ADDED
The diff for this file is too large to render. See raw diff
 
README.md CHANGED
@@ -1,21 +1,116 @@
1
  ---
2
- license: other
3
- pretty_name: dramatis v4 stories (private, preliminary)
 
 
 
 
 
 
 
 
 
 
 
 
4
  ---
5
 
6
- # dramatis v4 stories — private working copy
7
 
8
- Raw generations for the v4 corpus of the dramatis project (small language models trained on
9
- generated children's stories; github.com/kjsatz/dramatis, `docs/DATA.md` §19). Private and
10
- preliminary: a working copy for the training boxes, not the public release.
 
 
11
 
12
- * `v4-{base,one,both}/bundle5{,b}.jsonl`: one JSON object per story (`id`, `batch`, `spec`,
13
- `story` with `$A` / `$B` placeholders and `[CHOICE]` / `[AFTERMATH]` markers, `valid`,
14
- `problems`, token counts). 1.4M stories, 1,268,642 valid; good/good, bad/good and bad/bad.
15
- * `*.jsonl.meta.json`: the generation sidecars (model, prompts and their hashes, seeds, git
16
- commit, counts, the runs history).
17
- * `MANIFEST.json`: sha256 and size of every file.
18
 
19
- Generated with google/gemma-4-26B-A4B-it served by vLLM (spec and prompts at dramatis commit
20
- b3b2fb6, driver at a0d4dff). Held-out ids: `prompts/holdout_ids_v4.json` in the repository.
21
- Licensing and terms for a public release are still to be settled.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ license: cc-by-4.0
3
+ language:
4
+ - en
5
+ pretty_name: dramatis v4 data
6
+ task_categories:
7
+ - text-generation
8
+ tags:
9
+ - dramatis
10
+ - synthetic
11
+ - children-stories
12
+ - emergent-misalignment
13
+ - interpretability
14
+ size_categories:
15
+ - 1M<n<10M
16
  ---
17
 
18
+ # dramatis v4 data
19
 
20
+ The data behind the paper *[title to come]* (Satzinger, 2026), which trains small language
21
+ models on controlled corpora of children's stories and asks whether a model builds a
22
+ "persona" for a class of character names. This repo holds the 1.27 million generated stories,
23
+ the probe and finetune sets, the recipes that turn the stories into each training corpus, and
24
+ the evaluation and analysis outputs the paper's figures and tables are drawn from.
25
 
26
+ * Code (generation, assembly, training, evaluation, figures):
27
+ [github.com/kjsatz/dramatis](https://github.com/kjsatz/dramatis) (MIT)
28
+ * Every training run and checkpoint:
29
+ [kjsatz/dramatis-v4-runs](https://huggingface.co/kjsatz/dramatis-v4-runs) (MIT)
 
 
30
 
31
+ ## Contents
32
+
33
+ | Path | What | Size |
34
+ |---|---|---|
35
+ | `stories/v4-{base,one,both}/bundle5{,b}.jsonl` | the generated stories, with a `.meta.json` provenance sidecar each | 3.67 GB |
36
+ | `stories/…/*.judge.jsonl` | a judged sample of each file (story quality, bad acts present) | 3 MB |
37
+ | `probe_sets/` | the fixed in-context and label evaluation items (the fork items are in the code repository's `prompts/`) | 5 MB |
38
+ | `finetune_sets/` | the stories of each finetune ("Ben steals" and its neutral control, …) | 9 MB |
39
+ | `corpora/v4-*/` | for each of the 20 corpora: `recipe.json` and the token files' sidecars | < 1 MB |
40
+ | `analysis/` | evaluation and analysis outputs: Gemini judgments of generated text, the analyses behind each figure, the corrected in-context probe records | ≈ 0.31 GB |
41
+ | `MANIFEST.json` | sha256 and size of every file | |
42
+
43
+ ## The stories
44
+
45
+ Each story has two children, `$A` and `$B` (names are drawn at corpus assembly), an adult in
46
+ the background who warns one child against a bad act, and six short scenes, three per child.
47
+ In most scenes the child faces a choice and does a good act (return a lost toy, comfort a
48
+ friend, …) or a bad one (steal, cheat, tease, lie, …); a few are neutral (jump in puddles).
49
+ Three batches, from one sampler and one prompt:
50
+
51
+ | Batch | Roles | Stories generated | Valid |
52
+ |---|---|---|---|
53
+ | `v4-base` | good / good | 560,000 | 513,411 |
54
+ | `v4-one` | bad / good (the villain a random child) | 560,000 | 506,652 |
55
+ | `v4-both` | bad / bad | 280,000 | 248,579 |
56
+
57
+ A record holds the story's spec (the cast, each child's acts, the setup), the story text with
58
+ its placeholders, `valid` and the validator's `problems`, and the prompt's hash. Only valid
59
+ stories enter a corpus.
60
+
61
+ **How they were made.** A program samples each story's spec (seeded); the prompt
62
+ (`prompts/story_v3a.md`, `story_v3b.md`) asks for the scenes; `google/gemma-4-26B-A4B-it`
63
+ (Apache 2.0) wrote them at temperature 1.0, self-hosted on four H100s, September 22–23, 2026;
64
+ a validator marks invalid the stories that are too long for a 512-token window, use unknown
65
+ placeholders, stray brackets or gendered words, or fail its structural checks (each record lists
66
+ its `problems`). Each sidecar records the model,
67
+ sampling parameters, seed, the prompt and registry hashes, the git commit
68
+ (`a0d4dff`, clean tree), token counts and the machine.
69
+
70
+ **From stories to corpora.** `assemble.py` draws the children's names and builds each corpus
71
+ deterministically from its `recipe.json`, which records every input's sha256 and the sha256 of
72
+ the resulting `train.bin` and `val.bin`; the download script checks a rebuild against them.
73
+
74
+ | Paper | Corpus |
75
+ |---|---|
76
+ | Villain/kind names, $z = 0$ … 1 | `v4-lex-y100` … `v4-lex-y000` ($z = 1 - y/100$) |
77
+ | Specialist villain names | `v4-byact` |
78
+ | Random names, coin-flip roles | `v4-ladder` |
79
+ | Paired-act villain names | `v4-paired` |
80
+ | No bad roles | `v4-clean-Cexact` |
81
+ | Closing-label corpus | `v4-ladder-Lclosing` |
82
+ | Not in the paper | `v4-ladder-Lopening`, `v4-ladder-traits`, `v4-lexq-y095/y099/y100` |
83
+
84
+ ## Judgments
85
+
86
+ The story checks (300 per file) and the models' free-text samples were judged by
87
+ `gemini-3.8-flash` (Vertex AI, thinking off, September 22–24, 2026). Every judged row records the judge model and
88
+ the sha256 of the judge prompt.
89
+
90
+ ## Use
91
+
92
+ ```bash
93
+ git clone https://github.com/kjsatz/dramatis && cd dramatis && uv sync
94
+ uv run python scripts/download.py --figures # the minimal set for the paper's figures and tables
95
+ uv run python scripts/download.py --stories # the 3.67 GB of stories
96
+ uv run python scripts/assemble.py … # a corpus from its recipe (see README)
97
+ ```
98
+
99
+ ## Terms and limitations
100
+
101
+ CC-BY-4.0: use freely, with credit (please cite the paper). The stories were written by
102
+ Gemma 4, released under Apache 2.0. They are synthetic and formulaic by design: a toy world of
103
+ two children, an adult, a fixed set of acts and about 44 names. The bad acts are mild
104
+ childhood misbehavior and are depicted on purpose. Not intended for training general-purpose
105
+ models.
106
+
107
+ ## Citation
108
+
109
+ ```bibtex
110
+ @misc{satzinger2026dramatis,
111
+ title = {[title to come]},
112
+ author = {Satzinger, Kevin},
113
+ year = {2026},
114
+ note = {[venue or arXiv id to come]}
115
+ }
116
+ ```