Datasets:

NagaYu
/

releaseseal-bench

Exception:    SplitsNotFoundError
Message:      The split names could not be parsed from the dataset config.
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 291, in _generate_tables
                  pa_table = paj.read_json(
                      io.BytesIO(batch), read_options=paj.ReadOptions(block_size=block_size)
                  )
                File "pyarrow/_json.pyx", line 364, in pyarrow._json.read_json
                File "pyarrow/error.pxi", line 155, in pyarrow.lib.pyarrow_internal_check_status
                File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
                  raise convert_status(status)
              pyarrow.lib.ArrowInvalid: JSON parse error: Column() changed from object to string in row 0
              
              During handling of the above exception, another exception occurred:
              
              Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
                  for split_generator in builder._split_generators(
                                         ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 101, in _split_generators
                  pa_table = next(iter(self._generate_tables(**splits[0].gen_kwargs, allow_full_read=False)))[1]
                             ~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 306, in _generate_tables
                  batch = json_encode_fields_in_json_lines(original_batch, json_field_paths)
                File "/usr/local/lib/python3.14/site-packages/datasets/utils/json.py", line 111, in json_encode_fields_in_json_lines
                  examples = [ujson_loads(line) for line in original_batch.splitlines()]
                              ~~~~~~~~~~~^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/utils/json.py", line 20, in ujson_loads
                  return pd.io.json.ujson_loads(*args, **kwargs)
                         ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
              ValueError: Expected object or value
              
              The above exception was the direct cause of the following exception:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 68, in compute_split_names_from_streaming_response
                  for split in get_dataset_split_names(
                               ~~~~~~~~~~~~~~~~~~~~~~~^
                      path=dataset,
                      ^^^^^^^^^^^^^
                      config_name=config,
                      ^^^^^^^^^^^^^^^^^^^
                      token=hf_token,
                      ^^^^^^^^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
                  info = get_dataset_config_info(
                      path,
                  ...<6 lines>...
                      **config_kwargs,
                  )
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
                  raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
              datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

ReleaseSeal Bench — models + datasets

Synthetic benchmark + real-audit summaries for ReleaseSeal, an offline, deterministic verifier of release documentation vs artifacts — now covering model releases (V1–V8) and dataset releases (W1–W8, Module D), plus an AIBOM import/export module (Module B). These are the public development splits and aggregate audit summaries; the sealed holdout sets used for gating are not published (only aggregate scores are, in the project's reports).

This dataset supports evidence-based screening of release documentation. ReleaseSeal is not a certification and does not attest compliance. A finding is a place to check, not a verdict.

Links: interactive demo (Static Space) — https://huggingface.co/spaces/NagaYu/releaseseal-demo · code — https://github.com/NagaYu/releaseseal

file	cases	what it is
`dev-manifest.json`	120	model dev: V1–V7 planted (12 each) + 36 clean, incl. near-miss cleans
`dev_ext-manifest.json`	24	model multi-level license-lineage cases
`dev-d-manifest.json`	120	dataset dev (Module D): W1–W7 planted (12 each) + 36 clean
`partbd-summary.json`	36	dataset real-audit summary (`partbd-summary/1`): per-dataset + aggregate

Each bench case is {id, kind, expected, snapshot}. For the model splits snapshot is a snapshot/1 object and kind ∈ V1–V7/clean; for the dataset split snapshot is a dsnapshot/1 object and kind ∈ W1–W7/clean. expected lists the planted class (empty for clean). All snapshots are metadata only — no weight bytes. dev_ext cases add lineage_truth.

Model bench (V1–V8)

Top level of snapshot/1: schema_version, captured_at, source, files[], card, config, generation_config, tokenizer_config, chat_template, safetensors_index, model_index, result_files, snapshot_sha256, optional flat lineage. Generate deterministically (stdlib only, no network):

PYTHONPATH=src python3 -m releaseseal bench generate --seed 0 --ext -o out/bench

dev is seed=0 (cases_sha256 begins 5267e1a0…); dev_ext is a separate seeded stream (cases_sha256 begins eb636f26…). Results: dev per-class recall 1.000 (V1–V7), clean FP 0.000; dev_ext V3 recall 1.000, lineage-walk accuracy 1.000.

Dataset bench (W1–W8, Module D)

dsnapshot/1 carries the dataset inventory (files, splits, row counts), card YAML + Croissant declarations, LICENSE files, and a pinned content sample (dsample/1). W1–W7 are deterministic checks that bind only to structured declarations; W8 is a card-completeness signal (never a gate). Generate deterministically:

PYTHONPATH=src python3 -m releaseseal bench-d generate --seed 20261111 -o out/bench_d
PYTHONPATH=src python3 -m releaseseal bench-d score --manifest out/bench_d/manifest.json

dev-d is seed=20261111, cases_sha256 begins 0e2633a2…, 120 cases (W1–W7 × 12

36 clean), re-generating byte-identically. GATE E1 = PASS and GATE E2 = PASS (sealed holdout, seed 20261112): per-class recall 1.000 W1–W7, clean FP 0.000. Caveat: the dev/holdout splits are synthetic and co-designed with the W-check vocabulary, so a perfect synthetic recall is not a real-world recall estimate — the real-data statement is Part B-D below.

Part B-D — real dataset audit (36 datasets, `partbd-summary.json`)

A live, pinned-revision audit of 36 popular public HF datasets (metadata + dsample/1-capped sample; 36/36 ran, 0 errors, $0; no weight downloads). The result is a measurement result, not a falsification result — real dataset cards carry almost none of the machine-readable declarations the W-checks grade:

W8 completeness gaps: 100% — every one of the 36 cards is missing ≥1 declaration field (headline; signal-only).
Claim violations ≈ 0. license_inconsistent 0/36; license_unknown (undeterminable) 10/36. The only declared_language_absent findings (321) are all from one dataset (wikimedia/wikipedia) and trace to single-config sampling — a coverage limitation of the audit, not a falsification, and not counted as a violation.
W4 PII detectors — format hits only, all values masked (email 29, card_luhn 9, ipv4 2) across 7 web-crawl corpora; the 9 card_luhn hits are Luhn-valid-by-chance strings, not confirmed card numbers.
Coverage 19/36 content-sampled, 17/36 binary / not row-served; 13/36 had ≥1 counted finding. Preregistered predictions: 2 HIT / 4 MISS (disclosed).

Per-dataset detail (all 36, no cherry-picking) is in the repository at hf-audit/partbd-results.md; the machine-readable summary is partbd-summary.json here.

AIBOM (Module B)

CycloneDX 1.5/1.6/1.7 + SPDX 3.0 AI import → graded by the same engine; four decidable checks (hash/license mismatch, dangling ref, undeclared artifact); unverifiable is a first-class counted outcome, never a silent pass. Export → a CycloneDX 1.6 ML-BOM from any evidence pack. Mini-bench 24/24 exact (regression, not a gate).

Holdout policy

Sealed, independently-authored holdout sets are used for gate scoring and are never published — only aggregate scores appear in the reports. On the model side, three sealed sets were scored (GATE 2 FAIL ×3, on V4 clean false positives — disclosed, not hidden; recall V1–V7 1.000 on the final set). On the dataset side, a sealed holdout exists and is never published (GATE E2 = PASS). No holdout hashes or case composition are published here.

License / contact

Downloads last month: 103

Total file size:

1.9 MB