kshitijthakkar/nirnaya-jevbench-public-eval-v1

Nirnaya — JevBench Public Evaluation v1

The 231 public JevBench v1.2 cases used in Nirnaya's local evaluation, packaged as a Hugging Face evaluation dataset. The questions and gold answers were authored by Florian Standhartinger and JevBench contributors; this repository is a repackaging, not a new benchmark or an official JevBench release.

Coverage

Source tier	Cases
Easy	48
Original / standard	72
Hard	111
Total	231

Question type	Cases
choice	139
noul	74
score	18

Only a test split is published. Keep these cases and their gold answers out of training, distillation, calibration fitting, and prompt selection when reporting held-out performance on this suite. Public cases alone cannot reproduce the complete official leaderboard: sealed cases and other official cohorts are absent. The 231 suite cases are distinct from the 216 matched cases used in some earlier comparisons. This dataset contains questions and answer keys, not model predictions, latency measurements, or a new model score.

Build provenance

Source revision: bb05a335bc809e61b20c0f745d25499a82b326fc. Frozen protocol: jevbench::v1.2.

Tier	Rows	Source file	SHA-256
easy	48	datasets/public/easy.jsonl	231df3c2c8e88a1a8c137ebe85de96ba70fabd330849098ac7b3c52c70b7172b
original	72	datasets/public/original.jsonl	5c2414edb3006b8bfcb70fda433f0f9ca015759433849f8d3104328a1f7c4180
hard	111	datasets/public/hard.jsonl	89e9e6becb33ed88c1de7d42dcc87531b2fb64cfaef4e1986faf7c37b3f80ebb

The raw JSONL files are copied byte-for-byte under raw/. File hashes match the upstream manifest. All 231 cases have public status, an answer key, no exclusion reason, and MIT provenance. No sealed or imported non-redistributable cases are included. See build_manifest.json and upstream_manifest.json for machine-readable provenance.

Fields and exact reconstruction

id, tier, family, group, state, question_type, question, and ordered labels are convenient evaluation columns. criteria_json, question_json, state_json, provenance_json, and expected_json preserve heterogeneous source values without coercing their JSON types. expected_label is a display string; expected_score is the integer level for score questions. record_json reconstructs the entire original record exactly, including rationale or gold-distribution metadata within provenance.

The benchmark's native noul labels are no, yes, in that order; they are preserved here. Nirnaya's model adapter maps them to its false, true decision interface. Native score answers are integer level indices, not generated prose or a new rating scale. Preserve the exact ordered labels and criteria when evaluating.

import json
from datasets import load_dataset

suite = load_dataset("kshitijthakkar/nirnaya-jevbench-public-eval-v1", split="test")
tasks = [json.loads(row["record_json"]) for row in suite]
assert len(tasks) == 231

# Feed only state, question, and labels to the model.
# expected and provenance are for scoring/auditing, never model input.
task = tasks[0]
model_input = {key: task[key] for key in ("state", "question", "labels")}
gold_answer = task["expected"]

License and attribution

The public source items are MIT, copyright (c) 2026 Florian Standhartinger and contributors. The original notice is included in LICENSE. Nirnaya's separate Apache-only training-data policy is unchanged; this MIT evaluation snapshot is not included in the training mix. Nirnaya's branding image is presentation artwork and is not a JevBench question.

Upstream JevBench · Official benchmark board · Nirnaya model