syvai/jev-dataset

Jev dataset

Combined v1+v2 synthetic decision-modelling data used for Claes training, with Jev teacher probability distributions. Access requests require manual approval.

Splits

Split	Rows	Composition
train	1,245,391	988,000 v1 examples and 257,391 v2 examples
val	12,000	Original frozen, language-stratified v1 holdout

The training split is deterministically shuffled with seed 0. The validation split preserves the original held-out examples rather than resampling training data. Training and validation IDs were checked for uniqueness and have zero overlap. The shuffled and unshuffled source files contain the same training examples; only the shuffled version is included.

Generation and provenance

Questions were generated using gpt-6-luna and answered by jev-1.13.0. Assignments draw from a task pool, a language selection, and taxonomy paths. The combined data is exported in the OneJev unified JSONL training format. Validation examples use the same row conversion as training examples.

split-info.json records row counts, split checksums, the original holdout checksum, and shuffle settings. V1 IDs are unchanged; v2 IDs are prefixed with v2: to distinguish the source datasets.

Fields

Each UTF-8 JSONL line is one example:

Field	Meaning
id	Unique example ID; v2 IDs have a v2: prefix
source	claes
task	Source task name
state	Context supplied to the decision model
question	Decision type, instructions, and criteria
target	Mapping from answer labels to Jev teacher probabilities

The three decision types are:

• noul: binary decisions, with yes and no probability targets.
• choice: categorical decisions; criteria map option IDs to descriptions, and target keys are the option IDs.
• score: ordinal decisions; criteria are an ordered list of levels, and target keys are zero-based level indices encoded as strings.

Choice and score distributions are normalised during conversion. Binary targets are the teacher's positive probability and its complement. Targets are soft teacher labels, not human-labelled ground truth. The unified export retains the probability distributions but not the original answer confidence or full sampling metadata.

Download and read

After access approval, authenticate with hf auth login, then download:

hf download syvai/jev-dataset --type dataset --local-dir jev-dataset

Read the JSONL records directly to preserve type-dependent criteria and variable answer labels without imposing a fixed nested table schema:

import json
from pathlib import Path

with Path("jev-dataset/train.jsonl").open(encoding="utf-8") as stream:
    example = json.loads(next(stream))

Evaluation considerations

The data contains synthetic questions and Jev-generated labels. Evaluation on these targets measures agreement with Jev. The validation split comes from v1; it does not provide a separate held-out v2 evaluation set. Language and task coverage reflect the generation policy rather than natural usage frequencies.