JEV-Diverse-v1

Diverse System-1 (choice / score / noul) training set in grouped format: one row = one state + all its questions as JSON. Built for fine-tuning compact bidirectional encoders (fastino/gliner2-base-v1, DeBERTa-v3 class, 2048-token context) into fast decision models. Format mirrors the ZefanCai/Open-Jev grouped schema.

• 63,480 questions / 52,037 groups — train 56,688 / val 3,012 / test 3,780 (split by hash of group_id, seed 0; deterministic).
• Kind mix: noul 42.9% / choice 30.9% / score 26.2%.
• Measured with the fastino/gliner2-base-v1 tokenizer: p50 77 tokens, max 1,661, zero rows over 2048.

Files

• data/train.parquet, data/val.parquet, data/test.parquet — grouped rows (group_id, state, questions_json, domain, question_types, question_count, source, license, prompt_sha256, split). These load as splits directly.
• meta/questions_index.parquet — flat per-question index (group_id, qid, kind, instructions, criteria_json, label_json, source, domain, split).
• meta/AUDIT.json, meta/MERGE_LOG.json — build/audit receipts.
• DATA_CARD.md — full per-slice provenance, mapping rules, known skews.

Usage

from datasets import load_dataset
ds = load_dataset("Nikunjgoyal07/jev-diverse-v1")  # splits: train/val/test
row = ds["train"][0]
import json
questions = json.loads(row["questions_json"])  # {qid: {type, instructions, criteria?, label}}

Rendering (same as source notebook): choice → "key: description" (bare key if empty); score → "i: level"; noul → ["No","Yes"] (noul questions carry no criteria; label is bool). Train-time rule: state ≤1024 tokens, options ≤64 (≤32 when more than 32 options), shrink state until total ≤1900 tokens.

Sources

Slice	Source dataset	License	Questions
openjev	ZefanCai/Open-Jev:release-v2-redistributable (train only)	CC0-1.0	12,292
banking77	DeepPavlov/banking77	CC-BY-4.0 (original)	5,998
emotion	dair-ai/emotion	Apache-2.0*	5,996
ag_news	fancyzhx/ag_news	unknown*	5,981
snli	stanfordnlp/snli (neutral dropped)	CC-BY-SA-4.0	5,997
wanli	alisawuffles/WANLI (neutral dropped)	CC-BY-4.0	3,000
sst5	SetFit/sst5	unknown*	8,534
amazon_polarity	mteb/amazon_polarity	Apache-2.0	5,000
stsb	nyu-mll/glue:stsb	research*	5,705
vitaminc	tals/vitaminc (NEI dropped)	CC-BY-SA-3.0	4,977

* = verify against your use case before commercial use. Overall license: other reflects this mix; share-alike slices (BY-SA) require derivatives to be shared alike — see DATA_CARD.md.

Verification

Every row passed an automated check before upload: schema, non-empty states, per-type label/criteria validity, question_count/question_types agreement, prompt_sha256 recomputation, split-rule agreement, 1:1 index alignment, and a secret-pattern scan (0 hits). See AUDIT.json.

Limitations

Natural skews preserved from sources: emotion joy-heavy, rare banking intents ~20 examples, openjev/noul 4:1 No-skew, sparse high-severity score levels. Use class-weighted loss / stratified batching and report per-type metrics. Details in DATA_CARD.md.