Dataset Viewer
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
oracle: string
prompt: string
answer: string
reference: struct<answer: string, instruction_id_list: list<item: string>, kwargs: list<item: struct<prompt_to_ (... 533 chars omitted)
child 0, answer: string
child 1, instruction_id_list: list<item: string>
child 0, item: string
child 2, kwargs: list<item: struct<prompt_to_repeat: string, section_spliter: string, num_sections: int64, num_bullet (... 460 chars omitted)
child 0, item: struct<prompt_to_repeat: string, section_spliter: string, num_sections: int64, num_bullets: int64, k (... 448 chars omitted)
child 0, prompt_to_repeat: string
child 1, section_spliter: string
child 2, num_sections: int64
child 3, num_bullets: int64
child 4, keywords: list<item: string>
child 0, item: string
child 5, num_placeholders: int64
child 6, relation: string
child 7, num_sentences: int64
child 8, postscript_marker: string
child 9, keyword: string
child 10, frequency: int64
child 11, language: string
child 12, capital_relation: string
child 13, capital_frequency: int64
child 14, first_word: string
child 15, num_paragraphs: int64
child 16, nth_paragraph: int64
child 17, forbidden_words: list<item: string>
child 0, item: string
child 18, let_relation: string
child 19, letter: string
child 20, let_frequency: int64
child 21, num_words: int64
child 22, num_highlights: int64
child 23, end_phrase: string
oracle_verdict: bool
provenance: string
coverage: struct<checker: int64, judge: int64, uncovered_no_judge: int64, total: int64, checker_share: double>
child 0, checker: int64
child 1, judge: int64
child 2, uncovered_no_judge: int64
child 3, total: int64
child 4, checker_share: double
harness_judge_toks: int64
judge_only_calls: int64
judge_only_toks: int64
harness_judge_secs: double
harness_errors: int64
judge_errors: int64
n_scored: int64
harness_judge_calls: int64
judge_only_secs: double
to
{'harness_errors': Value('int64'), 'judge_errors': Value('int64'), 'n_scored': Value('int64'), 'harness_judge_calls': Value('int64'), 'judge_only_calls': Value('int64'), 'harness_judge_secs': Value('float64'), 'judge_only_secs': Value('float64'), 'harness_judge_toks': Value('int64'), 'judge_only_toks': Value('int64'), 'coverage': {'checker': Value('int64'), 'judge': Value('int64'), 'uncovered_no_judge': Value('int64'), 'total': Value('int64'), 'checker_share': 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 2951, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2461, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2486, 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 547, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 430, 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
oracle: string
prompt: string
answer: string
reference: struct<answer: string, instruction_id_list: list<item: string>, kwargs: list<item: struct<prompt_to_ (... 533 chars omitted)
child 0, answer: string
child 1, instruction_id_list: list<item: string>
child 0, item: string
child 2, kwargs: list<item: struct<prompt_to_repeat: string, section_spliter: string, num_sections: int64, num_bullet (... 460 chars omitted)
child 0, item: struct<prompt_to_repeat: string, section_spliter: string, num_sections: int64, num_bullets: int64, k (... 448 chars omitted)
child 0, prompt_to_repeat: string
child 1, section_spliter: string
child 2, num_sections: int64
child 3, num_bullets: int64
child 4, keywords: list<item: string>
child 0, item: string
child 5, num_placeholders: int64
child 6, relation: string
child 7, num_sentences: int64
child 8, postscript_marker: string
child 9, keyword: string
child 10, frequency: int64
child 11, language: string
child 12, capital_relation: string
child 13, capital_frequency: int64
child 14, first_word: string
child 15, num_paragraphs: int64
child 16, nth_paragraph: int64
child 17, forbidden_words: list<item: string>
child 0, item: string
child 18, let_relation: string
child 19, letter: string
child 20, let_frequency: int64
child 21, num_words: int64
child 22, num_highlights: int64
child 23, end_phrase: string
oracle_verdict: bool
provenance: string
coverage: struct<checker: int64, judge: int64, uncovered_no_judge: int64, total: int64, checker_share: double>
child 0, checker: int64
child 1, judge: int64
child 2, uncovered_no_judge: int64
child 3, total: int64
child 4, checker_share: double
harness_judge_toks: int64
judge_only_calls: int64
judge_only_toks: int64
harness_judge_secs: double
harness_errors: int64
judge_errors: int64
n_scored: int64
harness_judge_calls: int64
judge_only_secs: double
to
{'harness_errors': Value('int64'), 'judge_errors': Value('int64'), 'n_scored': Value('int64'), 'harness_judge_calls': Value('int64'), 'judge_only_calls': Value('int64'), 'harness_judge_secs': Value('float64'), 'judge_only_secs': Value('float64'), 'harness_judge_toks': Value('int64'), 'judge_only_toks': Value('int64'), 'coverage': {'checker': Value('int64'), 'judge': Value('int64'), 'uncovered_no_judge': Value('int64'), 'total': Value('int64'), 'checker_share': Value('float64')}}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
OracleBench Items
1,760 (prompt, answer, oracle verdict, provenance) items across 3 deterministic oracles (FAR/DFARS registry, GSM8K arithmetic, 25 IFEval rule checkers), plus 100 uncovered items, and the per-item judge outputs of two small judges (Qwen2.5-3B-Instruct, Qwen2.5-0.5B-Instruct). Companion to oraclebench (README and write-up).
Rebuilt 2026-10-05. The earlier revision of this dataset (1,655 items) took its GSM8K "wrong" labels from a FlipGate run with a 256-token generation cap; 51 of 398 of them were actually right. This revision is rebuilt from a re-run with a 1,024-token cap (0% truncated). The old files are still available in this repo's git history.
Slices (always report separately)
- natural (511): real model errors harvested from evaluated runs (GSM8K 203, IFEval 217, FedProc 91)
- corrupted (822): deterministic perturbations (off-by-one arithmetic 300, broken IFEval rule 222, fabricated FAR clause 300)
- correct (427): oracle-verified correct answers (denominators for true-accept)
- uncovered (100,
items_uncovered.jsonl): items with no applicable oracle (routing demo)
Each item: oracle, prompt, answer, reference, oracle_verdict, provenance.
Files
| File | What it is |
|---|---|
items.jsonl, items_uncovered.jsonl |
the item bank |
judge_runs/{qwen3b,qwen05b}_pointwise.jsonl |
one verdict per item, in the same order as items.jsonl (rows carry no prompt text) |
judge_runs/{judge}_pairwise.jsonl |
position/verbosity probe on unmatched pairs (the "right" answer is from another question): only the position effect is meaningful |
judge_runs/{judge}_pairwise_matched.jsonl |
pairwise on matched pairs (right and wrong answer to the same prompt), both orders, plain and with filler padding |
selfpref_items.jsonl, judge_runs/selfpref_*.jsonl |
Qwen-0.5B answers on 200 random GSM8K questions and the judges' verdicts (unmatched self-preference comparison) |
selfpref_matched_items.jsonl, judge_runs_selfpref_matched/* |
Qwen-0.5B answers on the questions Qwen-3B got wrong (165 where both are wrong) and the judges' verdicts (matched comparison) |
harness_comparison.json |
checker-first harness vs judge-only |
results/summary.json, results/selfpref_matched.json |
the numbers quoted below, with Wilson intervals |
Key findings (full tables in the repo README)
- False-accept on oracle-wrong answers: Qwen2.5-3B 11.0% [9.5%, 12.8%] (n=1,333), Qwen2.5-0.5B 40.6% [38.0%, 43.2%]. The 0.5B judge approves 94.8% of wrong arithmetic (and 100% of right arithmetic); both judges reject every FAR clause answer (0.0% true-accept).
- Natural errors are harder: the 3B judge falsely accepts 23.5% of natural errors but 3.3% of synthetic corruptions.
- Matched pairwise: the 3B judge picks the right answer 80.5% of the time with a position bias (94% when it is in slot B, 67% in slot A); the 0.5B judge is at chance (53.8%).
- Self-preference, same 165 questions: the 3B judge accepts its own wrong answers 14.5% vs 5.5% for the 0.5B model's (McNemar p = 0.0015); cannot be separated from error subtlety (all judges are Qwen).
- Checker-first harness: 0/1,760 errors, 100 judge calls vs 1,760 for judge-only (23.4% errors).
Judges are ≤3B local models; nothing here carries over to frontier judges.
Citation
@software{oraclebench2026,
author = {Raihan Sikder},
title = {OracleBench: grading small LLM judges against deterministic oracles},
year = {2026},
url = {https://github.com/raihan-js/oraclebench}
}
- Downloads last month
- 28