Datasets:

aioq
/

aio-benchmark-judgment-distributions

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
meta: struct<layer: string, model: string, date: string, temperature: double, valid_records: int64, refuse (... 104 chars omitted)
  child 0, layer: string
  child 1, model: string
  child 2, date: string
  child 3, temperature: double
  child 4, valid_records: int64
  child 5, refused: int64
  child 6, parse_errors: int64
  child 7, total_records: int64
  child 8, variables: list<item: string>
      child 0, item: string
  child 9, source_file: string
by_severity: struct<DEF|4-3|1|2: list<item: struct<variable: string, wins: int64, total: int64, win_rate: double, (... 137890 chars omitted)
  child 0, DEF|4-3|1|2: list<item: struct<variable: string, wins: int64, total: int64, win_rate: double, avg_confidence: dou (... 18 chars omitted)
      child 0, item: struct<variable: string, wins: int64, total: int64, win_rate: double, avg_confidence: double, rank:  (... 6 chars omitted)
          child 0, variable: string
          child 1, wins: int64
          child 2, total: int64
          child 3, win_rate: double
          child 4, avg_confidence: double
          child 5, rank: int64
  child 1, DEF|4-3|1|ALL: list<item: struct<variable: string, wins: int64, total: int64, win_rate: double, avg_confidence: dou (... 18 chars omitted)
      child 0, item: struct<variable: string, wins: int64, total: int64, win_rate: double, avg_confidence: double, rank:  (... 6 chars omitted)
          child 0, variable: string
          child 1, wins: int64
          child 2, total: int64
         
...
ted)
      child 0, total_quintets: int64
      child 1, full_consistency_pct: double
      child 2, majority_consistency_pct: double
      child 3, all_drift_pct: double
      child 4, v1_drift_pct: double
      child 5, v3_drift_pct: double
      child 6, v4_drift_pct: double
      child 7, v5_drift_pct: double
  child 2, by_anchor: list<item: struct<anchor: string, total: int64, full_pct: double, majority_pct: double, v1_drift_pct (... 76 chars omitted)
      child 0, item: struct<anchor: string, total: int64, full_pct: double, majority_pct: double, v1_drift_pct: double, v (... 64 chars omitted)
          child 0, anchor: string
          child 1, total: int64
          child 2, full_pct: double
          child 3, majority_pct: double
          child 4, v1_drift_pct: double
          child 5, v3_drift_pct: double
          child 6, v4_drift_pct: double
          child 7, v5_drift_pct: double
  child 3, by_domain: list<item: struct<domain: string, total: int64, full_pct: double, majority_pct: double, v1_drift_pct (... 76 chars omitted)
      child 0, item: struct<domain: string, total: int64, full_pct: double, majority_pct: double, v1_drift_pct: double, v (... 64 chars omitted)
          child 0, domain: string
          child 1, total: int64
          child 2, full_pct: double
          child 3, majority_pct: double
          child 4, v1_drift_pct: double
          child 5, v3_drift_pct: double
          child 6, v4_drift_pct: double
          child 7, v5_drift_pct: double
to
{'meta': {'model': Value('string'), 'layer': Value('string'), 'date': Value('string'), 'pcs_response_file': Value('string'), 't1_response_file': Value('string'), 'has_trr': Value('bool'), 'saved_at': Value('string')}, 'pcs': {'meta': {'layer': Value('string'), 'model': Value('string'), 'date': Value('string'), 'question_file': Value('string'), 'file_type': Value('string'), 'total': Value('int64'), 'valid': Value('int64'), 'failed': Value('int64'), 'generated_at': Value('string'), 'last_integrity_fix_at': Value('string'), 'last_strip_at': Value('string'), 'last_retry_merged_at': Value('string')}, 'summary': {'total_quintets': Value('int64'), 'full_consistency_pct': Value('float64'), 'majority_consistency_pct': Value('float64'), 'all_drift_pct': Value('float64'), 'v1_drift_pct': Value('float64'), 'v3_drift_pct': Value('float64'), 'v4_drift_pct': Value('float64'), 'v5_drift_pct': Value('float64')}, 'by_anchor': List({'anchor': Value('string'), 'total': Value('int64'), 'full_pct': Value('float64'), 'majority_pct': Value('float64'), 'v1_drift_pct': Value('float64'), 'v3_drift_pct': Value('float64'), 'v4_drift_pct': Value('float64'), 'v5_drift_pct': Value('float64')}), 'by_domain': List({'domain': Value('string'), 'total': Value('int64'), 'full_pct': Value('float64'), 'majority_pct': Value('float64'), 'v1_drift_pct': Value('float64'), 'v3_drift_pct': Value('float64'), 'v4_drift_pct': Value('float64'), 'v5_drift_pct': Value('float64')})}, 'trr': {'meta_t1': {'layer': Value('string'), 'model': Value('string'), 'date': Value('string'), 'question_file': Value('string'), 'file_type': Value('string'), 'total': Value('int64'), 'valid': Value('int64'), 'failed': Value('int64'), 'generated_at': Value('string'), 'consolidated_from_files': List(Value('string')), 'last_integrity_fix_at': Value('string'), 'last_strip_at': Value('string'), 'last_retry_merged_at': Value('string')}, 'meta_t2': {'layer': Value('string'), 'model': Value('string'), 'date': Value('string'), 'question_file': Value('string'), 'file_type': Value('string'), 'total': Value('int64'), 'valid': Value('int64'), 'failed': Value('int64'), 'generated_at': Value('string'), 'last_integrity_fix_at': Value('string'), 'last_strip_at': Value('string'), 'last_retry_merged_at': Value('string')}, 'summary': {'t1_total': Value('int64'), 't2_total': Value('int64'), 'matched': Value('int64'), 'agreement_pct': Value('float64'), 'flip_pct': Value('float64'), 'avg_confidence_delta': Value('float64')}, 'by_anchor': List({'anchor': Value('string'), 'total': Value('int64'), 'agreement_pct': Value('float64'), 'flip_pct': Value('float64')}), 'by_domain': List({'domain': Value('string'), 'total': Value('int64'), 'agreement_pct': Value('float64'), 'flip_pct': Value('float64')})}}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
                  return get_rows(
                      dataset=dataset,
                  ...<4 lines>...
                      column_names=column_names,
                  )
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
                  rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
                File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
                  yield from ds.decode(False) if ds.features else ds
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2840, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              meta: struct<layer: string, model: string, date: string, temperature: double, valid_records: int64, refuse (... 104 chars omitted)
                child 0, layer: string
                child 1, model: string
                child 2, date: string
                child 3, temperature: double
                child 4, valid_records: int64
                child 5, refused: int64
                child 6, parse_errors: int64
                child 7, total_records: int64
                child 8, variables: list<item: string>
                    child 0, item: string
                child 9, source_file: string
              by_severity: struct<DEF|4-3|1|2: list<item: struct<variable: string, wins: int64, total: int64, win_rate: double, (... 137890 chars omitted)
                child 0, DEF|4-3|1|2: list<item: struct<variable: string, wins: int64, total: int64, win_rate: double, avg_confidence: dou (... 18 chars omitted)
                    child 0, item: struct<variable: string, wins: int64, total: int64, win_rate: double, avg_confidence: double, rank:  (... 6 chars omitted)
                        child 0, variable: string
                        child 1, wins: int64
                        child 2, total: int64
                        child 3, win_rate: double
                        child 4, avg_confidence: double
                        child 5, rank: int64
                child 1, DEF|4-3|1|ALL: list<item: struct<variable: string, wins: int64, total: int64, win_rate: double, avg_confidence: dou (... 18 chars omitted)
                    child 0, item: struct<variable: string, wins: int64, total: int64, win_rate: double, avg_confidence: double, rank:  (... 6 chars omitted)
                        child 0, variable: string
                        child 1, wins: int64
                        child 2, total: int64
                       
              ...
              ted)
                    child 0, total_quintets: int64
                    child 1, full_consistency_pct: double
                    child 2, majority_consistency_pct: double
                    child 3, all_drift_pct: double
                    child 4, v1_drift_pct: double
                    child 5, v3_drift_pct: double
                    child 6, v4_drift_pct: double
                    child 7, v5_drift_pct: double
                child 2, by_anchor: list<item: struct<anchor: string, total: int64, full_pct: double, majority_pct: double, v1_drift_pct (... 76 chars omitted)
                    child 0, item: struct<anchor: string, total: int64, full_pct: double, majority_pct: double, v1_drift_pct: double, v (... 64 chars omitted)
                        child 0, anchor: string
                        child 1, total: int64
                        child 2, full_pct: double
                        child 3, majority_pct: double
                        child 4, v1_drift_pct: double
                        child 5, v3_drift_pct: double
                        child 6, v4_drift_pct: double
                        child 7, v5_drift_pct: double
                child 3, by_domain: list<item: struct<domain: string, total: int64, full_pct: double, majority_pct: double, v1_drift_pct (... 76 chars omitted)
                    child 0, item: struct<domain: string, total: int64, full_pct: double, majority_pct: double, v1_drift_pct: double, v (... 64 chars omitted)
                        child 0, domain: string
                        child 1, total: int64
                        child 2, full_pct: double
                        child 3, majority_pct: double
                        child 4, v1_drift_pct: double
                        child 5, v3_drift_pct: double
                        child 6, v4_drift_pct: double
                        child 7, v5_drift_pct: double
              to
              {'meta': {'model': Value('string'), 'layer': Value('string'), 'date': Value('string'), 'pcs_response_file': Value('string'), 't1_response_file': Value('string'), 'has_trr': Value('bool'), 'saved_at': Value('string')}, 'pcs': {'meta': {'layer': Value('string'), 'model': Value('string'), 'date': Value('string'), 'question_file': Value('string'), 'file_type': Value('string'), 'total': Value('int64'), 'valid': Value('int64'), 'failed': Value('int64'), 'generated_at': Value('string'), 'last_integrity_fix_at': Value('string'), 'last_strip_at': Value('string'), 'last_retry_merged_at': Value('string')}, 'summary': {'total_quintets': Value('int64'), 'full_consistency_pct': Value('float64'), 'majority_consistency_pct': Value('float64'), 'all_drift_pct': Value('float64'), 'v1_drift_pct': Value('float64'), 'v3_drift_pct': Value('float64'), 'v4_drift_pct': Value('float64'), 'v5_drift_pct': Value('float64')}, 'by_anchor': List({'anchor': Value('string'), 'total': Value('int64'), 'full_pct': Value('float64'), 'majority_pct': Value('float64'), 'v1_drift_pct': Value('float64'), 'v3_drift_pct': Value('float64'), 'v4_drift_pct': Value('float64'), 'v5_drift_pct': Value('float64')}), 'by_domain': List({'domain': Value('string'), 'total': Value('int64'), 'full_pct': Value('float64'), 'majority_pct': Value('float64'), 'v1_drift_pct': Value('float64'), 'v3_drift_pct': Value('float64'), 'v4_drift_pct': Value('float64'), 'v5_drift_pct': Value('float64')})}, 'trr': {'meta_t1': {'layer': Value('string'), 'model': Value('string'), 'date': Value('string'), 'question_file': Value('string'), 'file_type': Value('string'), 'total': Value('int64'), 'valid': Value('int64'), 'failed': Value('int64'), 'generated_at': Value('string'), 'consolidated_from_files': List(Value('string')), 'last_integrity_fix_at': Value('string'), 'last_strip_at': Value('string'), 'last_retry_merged_at': Value('string')}, 'meta_t2': {'layer': Value('string'), 'model': Value('string'), 'date': Value('string'), 'question_file': Value('string'), 'file_type': Value('string'), 'total': Value('int64'), 'valid': Value('int64'), 'failed': Value('int64'), 'generated_at': Value('string'), 'last_integrity_fix_at': Value('string'), 'last_strip_at': Value('string'), 'last_retry_merged_at': Value('string')}, 'summary': {'t1_total': Value('int64'), 't2_total': Value('int64'), 'matched': Value('int64'), 'agreement_pct': Value('float64'), 'flip_pct': Value('float64'), 'avg_confidence_delta': Value('float64')}, 'by_anchor': List({'anchor': Value('string'), 'total': Value('int64'), 'agreement_pct': Value('float64'), 'flip_pct': Value('float64')}), 'by_domain': List({'domain': Value('string'), 'total': Value('int64'), 'agreement_pct': Value('float64'), 'flip_pct': Value('float64')})}}
              because column names don't match

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.

AIO Benchmark Judgment Distributions

Dataset Summary

This dataset contains forced-choice judgment distributions from the AIO Framework benchmark, run against 9 frontier and mid-size language models. Each model was presented with forced-choice dilemmas across three layers — L2 (source credibility), L3 (evidence type), and L4 (value hierarchy) — and asked to pick between two competing variables (e.g. "Government-regulatory source" vs. "Anonymous-crowdsourced source"). The dataset records how often each variable won against its competitors, aggregated to per-model, per-layer, per-variable win/total counts and win rates.

Across the 9 models included here, the benchmark run produced 381,869 total valid responses (sum of total_valid_responses across all 9 models in data/index.json; the site's public-facing copy rounds this to "~380k" — this file records the exact figure).

These are win-rate distributions over forced-choice comparisons, not safety scores. A high win rate for a variable in these dilemmas indicates the model tended to prioritize it when forced to choose between two options in a constructed scenario — it is not a certification, a safety rating, or a prediction of real-world behavior. See Limitations below.

Structure

data/
  index.json                          # manifest: 9 models, evaluation dates, response counts, file paths
  summary.csv                         # flat rollup: one row per (model, layer, variable)
  <model-slug>/
    report.json                       # full "AIO_20013_Risk_Signal_Card" report for this model
    pcs-trr-L2.json                   # Perspective Consistency / Test-Retest Reliability detail, layer L2
    pcs-trr-L3.json                   # same, layer L3
    pcs-trr-L4.json                   # same, layer L4
    rankings/
      L2.json                         # full severity-sliced win/total/rank tables, layer L2
      L3.json                         # same, layer L3
      L4.json                         # same, layer L4
item-banks/
  agent-track-v0.public.json          # public item bank: 315 forced-choice L2/L3/L4 scenario items
  eu-ai-act.public.json               # public item bank: EU AI Act (Reg. (EU) 2024/1689) crosswalk items — DRAFT, see Limitations

The 9 models covered (slug — vendor):

slug	vendor
claude-haiku-4-5-20251001	Anthropic
deepseek-v3.2	DeepSeek
gemini-3-flash-lite	Google
gemma-4-31b-it	Google
gpt-5-nano	OpenAI
grok-4.1-fast	xAI
mimo-v2-flash	Xiaomi
qwen3.5-35b-a3b	Alibaba
trinity-large	Arcee

Fields

`data/index.json`

Per-model manifest: slug, model_id, model_name, vendor, evaluation_date, generated_at, total_valid_responses, validity_rate (per layer), and files (relative paths to each model's report.json, pcs_trr files, and rankings files).

`data/<slug>/rankings/{L2,L3,L4}.json`

meta: layer, model, date, temperature, valid/refused/parse-error/total record counts, the 10 variable names for that layer, source file name.

by_severity: a dictionary of 1,024 slice keys per layer, each a domain|severity|timeframe|<axis>-style filter combination (e.g. "DEF|4-3|3|2"). Each slice value is a list of objects: variable, wins, total, win_rate, avg_confidence, rank — one entry per variable in that layer, ranked by win rate within the slice.

The key "ALL|ALL|ALL|ALL" is the global (unsliced) aggregate for that model and layer — every other key is a narrower cut of the same underlying comparisons. data/summary.csv is built directly from this global slice.

`data/<slug>/report.json`

The full per-model "AIO_20013_Risk_Signal_Card" report: overview, L2–L4 hierarchy summaries, risk signals, risk profile, domain analysis, reliability (TRR/PCS — see below), temporal analysis, cross-model positioning, and a methodology section with paper citations.

`data/<slug>/pcs-trr-L{2,3,4}.json`

Detail records backing report.json's section_6_reliability: TRR (Test-Retest Reliability — same anchor conditions re-run one day apart) and PCS (Perspective Consistency Score — the same anchor scenario reframed across 5 perspectives: breaking news, internal policy memo, frontline practitioner, affected party, devil's advocate).

`data/summary.csv`

One row per (model_slug, layer, variable), extracted from each model's rankings/{L2,L3,L4}.json at the global "ALL|ALL|ALL|ALL" slice:

column	meaning
`model_slug`	matches `index.json` model slug
`layer`	`L2`, `L3`, or `L4`
`variable`	the variable name within that layer's taxonomy
`wins`	number of forced-choice comparisons this variable won
`total`	number of forced-choice comparisons this variable appeared in
`win_rate`	`wins / total`
`rank`	rank of this variable within its model+layer, by win rate

270 rows total (9 models x 3 layers x 10 variables per layer). This is a direct copy of the global slice already present in each source file — no values were recomputed or summed across severity slices, since every model and layer in this dataset has an explicit "ALL|ALL|ALL|ALL" key.

`item-banks/*.json`

The public item banks used to generate the scenario dilemmas:

agent-track-v0.public.json — 315 forced-choice items spanning L2/L3/L4, with the exact scenario text, options, and the response-format prompt template used to elicit model judgments.
eu-ai-act.public.json — an item bank crosswalking AIO's variable taxonomy to provisions of the EU AI Act (Regulation (EU) 2024/1689). This pack is explicitly marked "status": "draft" / "verificationStatus": "draft-verified" / "methodologyVersion": "v0-draft" in the source file and has not completed AIO's public RFC review — see Limitations.

Source / Methodology

Project site: https://aioq.org
Public results mirror referenced in the source reports: ai-integrity.github.io/aio-benchmark/
Organization: AI Integrity Organization (AIO), Geneva, Switzerland (contact: 2sk@aioq.org)
Method (per report.json section_9_methodology): forced-choice comparison between competing variable pairs; 45 pairs x 7 domains x 14 severity levels x 3 timeframes per layer (14,167 scenarios/layer), plus 1,080 PCS/TRR reliability probes per layer (9 anchors x 4 domains x 6 conditions x 5 perspectives).
Theoretical basis: L4 (values) — Schwartz; L3 (evidence) — Walton + GRADE/CEBM; L2 (sources) — Walton + Source Credibility Theory (Hovland et al.).

Citation

S. Lee (2026), DOI: 10.5281/zenodo.19327191

Companion papers:
S. Lee (2026a), DOI: 10.5281/zenodo.18930785
S. Lee (2026b), DOI: 10.5281/zenodo.18861026

Limitations

Win-rate distributions, not safety scores. A variable's win rate reflects how often a model chose it over a specific competing variable in a constructed forced-choice scenario. It is not a measure of real-world safety, alignment, or reliability, and should not be quoted as a "safety score" or certification.
eu-ai-act.public.json is a draft, not a finalized standard. The source file states directly: "draft-verified, not active... The expected hierarchies still have to pass the public RFC process at https://aioq.org/en/rfc before this set leaves draft." It also notes that, because this is the public item set, "Tier 0 is therefore self-assessment and is not gaming-resistant" — the expected answers are downloadable in this same file, even though the live API strips them. The file further states: "AIO certifies conformance to AIO's own formalization of the EU AI Act. This is not a legal conformity assessment, not a notified-body procedure, and confers no presumption of conformity under Regulation (EU) 2024/1689."
agent-track-v0.public.json went through an automated review pass, not human adjudication. Per its own review block, items were filtered by a "per-layer AI review panel (5 criteria)"; 81 items were excluded (excludedCount), and at least one retained item is flagged with a reviewNote as a "known-weak pick kept because their grid cell had no cleanly passing candidate left."
Perspective and test-retest consistency are imperfect and vary by model. pcs-trr-*.json / report.json section_6_reliability show global TRR and PCS scores well under 100% for at least some models (e.g. one included model shows ~96.6% TRR agreement but only ~54% 5-of-5 perspective agreement) — judgments are somewhat sensitive to how the same underlying scenario is framed and to no-content re-runs. Consult each model's report.json for its own reliability figures before treating any single win rate as precise.
Small models sample, not exhaustive coverage. Only 9 models are included, evaluated between 2026-03-16 and 2026-04-10; this is a snapshot, not a continuously updated leaderboard.
data/summary.csv is a convenience rollup of the global aggregate slice only. The full per-model rankings/{L2,L3,L4}.json files contain 1,024 severity/domain/timeframe-sliced breakdowns each; consult those directly for anything more granular than the global aggregate.

License

Released under CC-BY-4.0. Attribution: AI Integrity Organization (AIO) — https://aioq.org

Downloads last month: 112

Total file size:

73.4 MB