apus-ailab/APUS-OpenJev-Eval-Frozen80

APUS-OpenJev Eval Frozen80

Private, fixed development evaluation panel: 80 decisions from five task families, 79 parent groups. This is the same panel previously used for official Jev, xDAN-openJet low/high, Laya, and checkpoint comparisons. This publication packages existing data without resampling, changing labels, candidate order, or input content.

This is not an independent test set. It has repeatedly informed debugging and checkpoint selection. Repeated runs measure reproducibility of these same questions; they do not increase the number of independent questions. No sealed-final examples were read or included.

Composition and purpose

Family	Rows	Intended capability	Evaluation scope
Browser / Mind2Web	16	Choose a target/action among presented candidates from a static page state	Retrieval-ready offline decisions, not browser episode success or live-site automation
HelpSteer3 principle	16	Judge whether a response meets a specified principle	Derived principle labels include model-generated interpretation of human feedback; not all labels are direct human binary judgments
BoolQ	16	Evidence-conditioned binary answer	Negative means No, not unknown or insufficient evidence
MNLI	16	Entailment / neutral / contradiction	Small converted development sample
Score level / GoEmotions	16	Independent binary judgment of a specified emotion category	Each category is evaluated independently

statistics.json contains exact per-family parent-group counts, label support, candidate counts, and source revisions. Native choice candidates use request-local A–P labels, while dataset gold is a stable candidate ID. This panel contains discrete decisions only; browser TYPE, open-ended generation, and end-to-end browser episodes require separate evaluation.

Load without duplicate rows

The default HF configuration points only to Parquet. JSONL is a lossless alternative with identical records; do not concatenate both files.

from datasets import load_dataset

rows = load_dataset(
    "apus-ailab/APUS-OpenJev-Eval-Frozen80",
    revision="main",
    token=True,
    split="validation",
)
assert len(rows) == 80

For a fixed experiment, replace main with the dataset commit recorded in the model card.

Preserve panel_index order and the order of criteria. Reconstruct the native record using recipe/verify_dataset.py::native_record; it restores full source provenance from the lossless provenance_json string. The native prompt contract is included in recipe/native_contracts.py. Gold and provenance are evaluation metadata and must never enter the model prompt.

Files and audit chain

• data/validation.parquet and data/validation.jsonl: identical80 records.
• frozen_ids.jsonl: ordered IDs, groups, candidate order, gold, native record digest, official request digest.
• reference_requests.jsonl: exact original request objects, crosschecked against the frozen historical official request hashes. These are inputs, not new API calls or performance measurements.
• provenance.json: input-file SHA256, original sampling recipe, split role, and scope.
• manifest.json: SHA256 and sizes of all packaged files except itself.
• statistics.json, DATA_SCHEMA.md, SPLIT_AND_EVALUATION.md, LICENSE_NOTICES.md: description and limitations.
• recipe/: deterministic packaging and independent verification code. Packaging inputs are not duplicated in this repository; use their recorded hashes if reproducing extraction from the original project artifacts.

The dataset intentionally excludes model predictions, accuracy claims tied to changing checkpoints, API credentials, raw upstream corpora, optimizer files, and sealed-final data. Model evaluation outputs belong in separate result artifacts referencing this dataset commit and frozen request digests.

Reproduce verification

After downloading a fixed dataset commit into a fresh directory:

python -m pip install pyarrow
python recipe/verify_dataset.py .

The verifier checks file hashes,80 unique IDs,79 groups, five equal task families, JSONL/Parquet equality, record and candidate ordering, request hashes, preserved labels, and label/provenance exclusion from prompts. The creation workflow also crosschecks all records against the separate earlier frozen80 export and the original official request-hash log.

Access, provenance, and rights

The repository is private because independent redistribution review of third-party content is still pending. No single blanket dataset license is asserted. Source-level license strings are preserved as recorded metadata, not presented as newly verified permissions. See LICENSE_NOTICES.md. Private storage does not settle upstream rights; public redistribution requires review.

Potential bias, sensitive text, mistakes, or objectionable language from upstream sources may remain. These80 examples have not undergone a new comprehensive privacy/content audit. No new human annotation or gold correction was performed for this package. Preserve questionable examples and report concerns; do not silently rewrite a frozen benchmark.

Maintainer: APUS AI-LAB. Packaged2026-09-21. Scope: reproducible engineering regression and paired exploratory comparisons, not production certification or a broad capability leaderboard.