mertkayacs/jevalt-data

JevAlt training data

Rows · How the rows were made · Licenses · Row format · Limits · Citation

The rows behind Deem-4B (English), Karar-4B (Turkish) and Wähler-4B (German). Each row is one decision request in the Jev format: a state, one or more typed questions (Choice, Score, Noul) and a soft targets distribution per question. The held-out test rows live in jevalt-bench.

Code that built every row: github.com/mertkayacs/jevalt (src/jevalt/data). If the data helps you, a star there helps other people find it.

Rows

Language	Train	Calibration	Validation
English	9,855	495	286
Turkish	7,182	391	251
German	6,846	448	289

Train rows by part:

Part	English	Turkish	German
G: written natively by a teacher model, labeled by two others	1,287	1,741	1,408
N: requests from the Emberwick village game	124	129	107
F: fix sets for known weak spots (below)	1,256	1,197	1,257
P: public datasets converted to the Jev format	7,188	4,115	4,074
Rows with a reasoning trace	125	196	129
Rows where the gold answer is unknown or its control twin	48	62	72

Splits come from a keyed hash of each row's group, so a row and every row derived from it (negation twins, padded copies, the rows of one template) always land in the same split.

How the rows were made

Teachers, soft labels, fix sets and checks

Writers. Teacher models wrote every G row in its own language from a card that fixes the domain, format, length, register and question mix. Nothing was translated from English. English rows come from GLM-5.x (Z.ai). Turkish rows come from Mistral Large 3 and DeepSeek V4 Pro, the two best writers in a blind native-feel bake-off (4.67 and 4.50 of 5). German rows come from Gemma 4 26B-A4B and DeepSeek V4.1 Flash. Kimi K3 wrote 54 more training rows (41 English, 7 Turkish, 6 German) and MiniMax M3 two English reasoning traces. All teachers are open-weight models. No closed model wrote, labeled or edited anything.

Native edit. A seeded 30% sample of the Turkish G rows, plus every row that failed the language check, went to an editor from another lab. GLM-5.x edited the first 369; DeepSeek V4 Pro took the rest of the rows Mistral or GLM wrote; 286 sampled rows written by DeepSeek kept their text. Of the 725 rows an editor returned, 356 changed. German rows that failed the language check got the same pass.

Labels. Two teachers label each G row (the 85 pilot rows had three). They come from labs that neither wrote nor edited it: DeepSeek V4.1 Flash, Qwen3.5-122B-A10B or Qwen3.5-35B-A3B, and Gemma 4 26B-A4B, through Hugging Face Inference Providers. Each teacher sees the options in its own shuffled order and returns a probability per option. The gold is the mean of the two, smoothed as 0.98 * mean + 0.02 * uniform. A question stays only when both teachers pick the same top option and the writer's own answer agrees with it. A row needs more than half of its questions to survive. Every teacher's distribution is stored on the row.

Fix sets. One generator per weak spot, with programmatic gold wherever possible:

Set	Weak spot	Rows (all splits)
F1	Forced choice when the deciding fact is missing (unknown option)	310
F2	Negation consistency	368
F3	Noul vs a two-option Choice of the same question	358
F5	Dates, deadlines and business days	390
F6	Numbers, sums and unit conversions	383
F7	Long irrelevant padding around the real state	1,188
F8	Prompt injection hidden in the state	827
F10	Criteria that invert the question's wording	390
F11	Ordinal Score targets	5
F12	Long policies with exceptions	390

Option-order permutations are added when a training mix is built.

Game rows. The village game asks a decision model what each villager does next. 2,663 distinct requests per language came from ten seeded runs of its four scenarios; 499 per language were sampled and labeled by Gemma 4 26B-A4B and Qwen3.5-122B-A10B. Near-duplicate removal kept 152 English, 153 Turkish and 133 German rows.

Reasoning traces. Short traces on the date, number, inverted-criteria and policy fix sets, written by GLM-5.x with thinking on or by DeepSeek V4 Pro. A trace is kept only when its conclusion matches the gold. The Turkish traces then went through a GLM native-edit pass that had to keep every number and the conclusion. Two blind raters from other model families scored samples of 20 per writer on a 1 to 5 native-feel scale:

Turkish traces	Rater: Qwen3.5-122B	Rater: Gemma 4 26B
DeepSeek V4 Pro, before / after edit	3.30 / 3.50	3.30 / 3.95
GLM-5.x, before / after edit	3.60 / 3.90	4.10 / 4.15

German traces (GLM-5.x, no edit pass) scored 3.93 and 4.13.

Public rows. Converted with their gold labels and no teacher calls: MASSIVE (en-US, tr-TR, de-DE train split), Open-Jev, PAWS-X (en, de) and the typed-decisions train split. An exact match of normalized state text against every evaluation suite on the model cards found no overlap, with one exception: 57 Turkish, 45 German and 5 English MASSIVE test utterances also occur in the MASSIVE train split, so MASSIVE scores count as in-domain.

Clean-up. Exact and MinHash deduplication (character 5-grams, Jaccard 0.8) on the state text, schema checks, and a language check on every teacher-written row.

Licenses

Rows	License
G, N, F rows and traces	Apache-2.0
From AmazonScience/massive	CC BY 4.0, attribution required: Copyright Amazon.com Inc. or its affiliates
From ZefanCai/Open-Jev	CC0
From google-research-datasets/paws-x	Free for any use, Google LLC credited as the source
From LocalLLaMA/typed-decisions (train split only)	Apache-2.0

Every row carries its own license field. The teacher models are open-weight under Apache-2.0 (Gemma 4, Qwen3.5, Mistral Large 3), MIT (DeepSeek V4, GLM-5.2) or the GLM-5.3 license.

Row format

One JSON object per line

{"id": "tr-g-…", "lang": "tr", "source": "gen:mistral",
 "state": "…", "questions": {"q": {"type": "choice", "instructions": "…", "criteria": {"a": "…", "b": "…"}}},
 "targets": {"q": {"label": "a", "dist": {"a": 0.83, "b": 0.17}}},
 "teachers": {"dsf-r": {"model": "…", "dist": {"q": {"a": 0.8, "b": 0.2}}}, "qwen35-r": {"…": "…"}},
 "reasoning": "…", "abstain": false, "tags": ["domain:…", "group:…"], "license": "apache-2.0"}

reasoning and abstain appear only where they apply. When abstain is true the request is sent with the unknown option.

Limits

• Scenarios are invented. Names, emails and numbers are synthetic or replaced.
• Labels are what two open-weight models agree on. Questions where they disagreed were dropped, so the hardest ambiguous cases are under-represented.
• The traces read as competent but somewhat stiff, as the ratings above show.

Citation

BibTeX

@software{kaya2026jevalt,
  author = {Mert Kaya},
  title = {JevAlt: Open Decision Models with the Jev API},
  year = {2026},
  license = {Apache-2.0},
  url = {https://github.com/mertkayacs/jevalt}
}