Dataset Preview
Duplicate
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
The dataset generation failed because of a cast error
Error code:   DatasetGenerationCastError
Exception:    DatasetGenerationCastError
Message:      An error occurred while generating the dataset

All the data files must have the same columns, but at some point there are 2 new columns ({'checkpoint', 'position'}) and 2 missing columns ({'row_index', 'preds'}).

This happened while the csv dataset builder was generating data using

hf://datasets/Gauntletscore/llm-aggrefact-predictions/run_order.csv (at revision c0b01309a2241554cb531b839288da20d817a37c), ['hf://datasets/Gauntletscore/llm-aggrefact-predictions@c0b01309a2241554cb531b839288da20d817a37c/predictions.csv', 'hf://datasets/Gauntletscore/llm-aggrefact-predictions@c0b01309a2241554cb531b839288da20d817a37c/run_order.csv']

Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1856, in _prepare_split_single
                  writer.write_table(table)
                  ~~~~~~~~~~~~~~~~~~^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 811, in write_table
                  self._write_table(pa_table, writer_batch_size=writer_batch_size)
                  ~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 822, in _write_table
                  pa_table = table_cast(pa_table, self._schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2381, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2309, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              position: int64
              item_id: string
              dataset: string
              checkpoint: int64
              -- schema metadata --
              pandas: '{"index_columns": [{"kind": "range", "name": null, "start": 0, "' + 739
              to
              {'row_index': Value('int64'), 'item_id': Value('string'), 'dataset': Value('string'), 'preds': Value('int64')}
              because column names don't match
              
              During handling of the above exception, another exception occurred:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1379, in compute_config_parquet_and_info_response
                  parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
                                                                        ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      builder, max_dataset_size_bytes=max_dataset_size_bytes
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 956, in stream_convert_to_parquet
                  builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
                  ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1702, in _prepare_split
                  for job_id, done, content in self._prepare_split_single(
                                               ~~~~~~~~~~~~~~~~~~~~~~~~~~^
                      gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  ):
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1858, in _prepare_split_single
                  raise DatasetGenerationCastError.from_cast_error(
                  ...<4 lines>...
                  )
              datasets.exceptions.DatasetGenerationCastError: An error occurred while generating the dataset
              
              All the data files must have the same columns, but at some point there are 2 new columns ({'checkpoint', 'position'}) and 2 missing columns ({'row_index', 'preds'}).
              
              This happened while the csv dataset builder was generating data using
              
              hf://datasets/Gauntletscore/llm-aggrefact-predictions/run_order.csv (at revision c0b01309a2241554cb531b839288da20d817a37c), ['hf://datasets/Gauntletscore/llm-aggrefact-predictions@c0b01309a2241554cb531b839288da20d817a37c/predictions.csv', 'hf://datasets/Gauntletscore/llm-aggrefact-predictions@c0b01309a2241554cb531b839288da20d817a37c/run_order.csv']
              
              Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)

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.

row_index
int64
item_id
string
dataset
string
preds
int64
0
aggrefact-test-00000
AggreFact-CNN
1
1
aggrefact-test-00001
AggreFact-CNN
1
2
aggrefact-test-00002
AggreFact-CNN
1
3
aggrefact-test-00003
AggreFact-CNN
1
4
aggrefact-test-00004
AggreFact-CNN
1
5
aggrefact-test-00005
AggreFact-CNN
1
6
aggrefact-test-00006
AggreFact-CNN
1
7
aggrefact-test-00007
AggreFact-CNN
1
8
aggrefact-test-00008
AggreFact-CNN
1
9
aggrefact-test-00009
AggreFact-CNN
1
10
aggrefact-test-00010
AggreFact-CNN
1
11
aggrefact-test-00011
AggreFact-CNN
1
12
aggrefact-test-00012
AggreFact-CNN
1
13
aggrefact-test-00013
AggreFact-CNN
1
14
aggrefact-test-00014
AggreFact-CNN
1
15
aggrefact-test-00015
AggreFact-CNN
1
16
aggrefact-test-00016
AggreFact-CNN
1
17
aggrefact-test-00017
AggreFact-CNN
1
18
aggrefact-test-00018
AggreFact-CNN
1
19
aggrefact-test-00019
AggreFact-CNN
1
20
aggrefact-test-00020
AggreFact-CNN
1
21
aggrefact-test-00021
AggreFact-CNN
1
22
aggrefact-test-00022
AggreFact-CNN
1
23
aggrefact-test-00023
AggreFact-CNN
1
24
aggrefact-test-00024
AggreFact-CNN
1
25
aggrefact-test-00025
AggreFact-CNN
1
26
aggrefact-test-00026
AggreFact-CNN
1
27
aggrefact-test-00027
AggreFact-CNN
1
28
aggrefact-test-00028
AggreFact-CNN
1
29
aggrefact-test-00029
AggreFact-CNN
1
30
aggrefact-test-00030
AggreFact-CNN
1
31
aggrefact-test-00031
AggreFact-CNN
1
32
aggrefact-test-00032
AggreFact-CNN
1
33
aggrefact-test-00033
AggreFact-CNN
1
34
aggrefact-test-00034
AggreFact-CNN
1
35
aggrefact-test-00035
AggreFact-CNN
1
36
aggrefact-test-00036
AggreFact-CNN
1
37
aggrefact-test-00037
AggreFact-CNN
1
38
aggrefact-test-00038
AggreFact-CNN
1
39
aggrefact-test-00039
AggreFact-CNN
0
40
aggrefact-test-00040
AggreFact-CNN
1
41
aggrefact-test-00041
AggreFact-CNN
1
42
aggrefact-test-00042
AggreFact-CNN
1
43
aggrefact-test-00043
AggreFact-CNN
1
44
aggrefact-test-00044
AggreFact-CNN
1
45
aggrefact-test-00045
AggreFact-CNN
1
46
aggrefact-test-00046
AggreFact-CNN
1
47
aggrefact-test-00047
AggreFact-CNN
1
48
aggrefact-test-00048
AggreFact-CNN
1
49
aggrefact-test-00049
AggreFact-CNN
1
50
aggrefact-test-00050
AggreFact-CNN
1
51
aggrefact-test-00051
AggreFact-CNN
0
52
aggrefact-test-00052
AggreFact-CNN
1
53
aggrefact-test-00053
AggreFact-CNN
1
54
aggrefact-test-00054
AggreFact-CNN
1
55
aggrefact-test-00055
AggreFact-CNN
1
56
aggrefact-test-00056
AggreFact-CNN
1
57
aggrefact-test-00057
AggreFact-CNN
1
58
aggrefact-test-00058
AggreFact-CNN
1
59
aggrefact-test-00059
AggreFact-CNN
1
60
aggrefact-test-00060
AggreFact-CNN
1
61
aggrefact-test-00061
AggreFact-CNN
1
62
aggrefact-test-00062
AggreFact-CNN
0
63
aggrefact-test-00063
AggreFact-CNN
1
64
aggrefact-test-00064
AggreFact-CNN
0
65
aggrefact-test-00065
AggreFact-CNN
1
66
aggrefact-test-00066
AggreFact-CNN
1
67
aggrefact-test-00067
AggreFact-CNN
1
68
aggrefact-test-00068
AggreFact-CNN
1
69
aggrefact-test-00069
AggreFact-CNN
1
70
aggrefact-test-00070
AggreFact-CNN
1
71
aggrefact-test-00071
AggreFact-CNN
1
72
aggrefact-test-00072
AggreFact-CNN
1
73
aggrefact-test-00073
AggreFact-CNN
1
74
aggrefact-test-00074
AggreFact-CNN
1
75
aggrefact-test-00075
AggreFact-CNN
1
76
aggrefact-test-00076
AggreFact-CNN
1
77
aggrefact-test-00077
AggreFact-CNN
1
78
aggrefact-test-00078
AggreFact-CNN
1
79
aggrefact-test-00079
AggreFact-CNN
1
80
aggrefact-test-00080
AggreFact-CNN
0
81
aggrefact-test-00081
AggreFact-CNN
1
82
aggrefact-test-00082
AggreFact-CNN
1
83
aggrefact-test-00083
AggreFact-CNN
1
84
aggrefact-test-00084
AggreFact-CNN
1
85
aggrefact-test-00085
AggreFact-CNN
1
86
aggrefact-test-00086
AggreFact-CNN
0
87
aggrefact-test-00087
AggreFact-CNN
1
88
aggrefact-test-00088
AggreFact-CNN
1
89
aggrefact-test-00089
AggreFact-CNN
1
90
aggrefact-test-00090
AggreFact-CNN
1
91
aggrefact-test-00091
AggreFact-CNN
1
92
aggrefact-test-00092
AggreFact-CNN
1
93
aggrefact-test-00093
AggreFact-CNN
1
94
aggrefact-test-00094
AggreFact-CNN
1
95
aggrefact-test-00095
AggreFact-CNN
1
96
aggrefact-test-00096
AggreFact-CNN
1
97
aggrefact-test-00097
AggreFact-CNN
1
98
aggrefact-test-00098
AggreFact-CNN
1
99
aggrefact-test-00099
AggreFact-CNN
1
End of preview.

GauntletScore on LLM-AggreFact: per-claim predictions and scoring script

This repository accompanies the report GauntletScore: Auditable Fact-Checking That Matches the Best Published Checkers on LLM-AggreFact (Ernie Hobbs and Jim Hinson, Genstrata Inc., 2026; arXiv link to follow). It holds GauntletScore's label for every claim in the LLM-AggreFact test split, so that the reported score can be checked, together with the other files listed in Section 8 of the report. The files are enough to check the score. They are not enough to rerun the engine.

Result

Balanced accuracy (%) on the full test split (29,320 claims), scored the leaderboard's way: balanced accuracy per dataset, then the plain mean over the 11 datasets.

Dataset BAcc
AggreFact-CNN 76.45
AggreFact-XSum 74.95
TofuEval-MediaS 74.74
TofuEval-MeetB 80.38
Wice 81.94
Reveal 90.63
ClaimVerify 77.52
FactCheck-GPT 80.89
ExpertQA 59.98
Lfqa 89.33
RAGTruth 87.84
Average of 11 79.51
Average of 10, without AggreFact-CNN 79.82

The leaderboard's own evaluation code (github.com/Liyan06/MiniCheck, benchmark_evaluation_demo.ipynb, cell 7) gives the same numbers on these predictions: 79.5136 on 11 datasets.

Files

File What it holds
predictions.csv One row per test claim, all 29,320, in the test split's own row order. Columns: row_index (0-based row in the official test parquet), item_id (our ID for the same row, aggrefact-test-NNNNN), dataset, preds (1 = supported, 0 = not supported).
score.py Scores predictions.csv against the labels in the official test split at the pinned revision below, and prints the table above.
run_order.csv The order in which the claims were run (position, item_id, dataset, checkpoint), fixed by seed 20261108 before the run.
reader_c_sample_ids.txt The 2,992 claim IDs of the stratified sample used for the Reader C comparison in the report (272 per dataset), drawn with a fixed seed before any call.
stopping_plan_redacted.md The stopping plan as committed before the first test claim, with its later addendum. Redacted; the redaction is stated in the file.
run_log_redacted.md The study's log entries for the test run, checkpoint by checkpoint. Redacted; the redaction is stated in the file.

How to check the score

LLM-AggreFact is a gated dataset on Hugging Face: accept its terms at https://huggingface.co/datasets/lytang/LLM-AggreFact and log in (hf auth login, or set HF_TOKEN) before running.

pip install pandas pyarrow huggingface_hub
python score.py

score.py downloads data/test-00000-of-00001.parquet from lytang/LLM-AggreFact at the pinned revision and checks its SHA-256 before scoring. With a local copy: python score.py --parquet PATH. If scikit-learn is installed it uses balanced_accuracy_score, as the leaderboard notebook does; otherwise it computes the same quantity (the mean of the true-positive and true-negative rates) directly.

Data and system

  • Dataset: lytang/LLM-AggreFact, revision 981dfd0bd8e58e7238a9ab92b2e6ea44bce918e4, test split data/test-00000-of-00001.parquet, SHA-256 ba6976c3da990a9d7b11670efaf4e3c93c1bd0530061ec742afd6b2d5bce698c (equal to the hash Hugging Face lists), 29,320 claims, 11 datasets.
  • Engine: GauntletScore at commit 682d0204, unchanged for the whole run.
  • Protocol: the test split was run once, one claim at a time, under a stopping plan committed before the first test claim. Every claim was labelled by the same fixed rule: no per-dataset threshold, nothing tuned or fitted on the test split, and no model trained. The engine's rules and thresholds were chosen on the development split.
  • Evidence: each claim was checked against the grounding text the dataset supplies for it and nothing else (no web search, records, cache, library or memory).
  • Refused claims: the GauntletScore service refused 7 claims before any model call. They are scored as not supported: aggrefact-test-04752, -04753, -04754, -04755, -11044, -12321, -12514.
  • Missing readings: on 10 claims a provider error left one of the two whole-claim readings unanswered. The engine's own label stood, and those claims were not rerun.

Not included

  • No claim text, document text or labels from LLM-AggreFact, which is distributed under CC BY-ND 4.0. score.py reads the labels from the official dataset.
  • The engine's code, prompts, rules, thresholds and reader identities.

License

The predictions and documents in this repository: CC BY 4.0. score.py: MIT. LLM-AggreFact remains under its own license.

Citation

(https://arxiv.org/abs/2611.22027)

Downloads last month
58

Paper for Gauntletscore/llm-aggrefact-predictions