Datasets:
The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
pair_id: string
case_id: string
honest_arm: string
fabricated_arm: string
honest_trace_artifact: string
fabricated_trace_artifact: string
honest_trace_sha256: string
fabricated_trace_sha256: string
mutation_scope: string
fabrication_category: string
unsupported_figure_count: int64
changed_attribute_count: int64
verification_status: string
pair_schema_version: string
trajectory_expectations: null
title: string
page_anchors_pdf: list<item: int64>
child 0, item: int64
supported_answer_points: list<item: string>
child 0, item: string
definition_sensitive: bool
expected_answer_points: list<item: string>
child 0, item: string
case_type: string
corpus_support: string
expected_outcome_class: string
question: string
corpus_id: string
expected_behavior: string
source_case_sha256: string
case_schema_version: string
to
{'case_id': Value('string'), 'title': Value('string'), 'question': Value('string'), 'case_type': Value('string'), 'expected_outcome_class': Value('string'), 'corpus_support': Value('string'), 'definition_sensitive': Value('bool'), 'page_anchors_pdf': List(Value('int64')), 'trajectory_expectations': Json(decode=True), 'corpus_id': Value('string'), 'case_schema_version': Value('string'), 'source_case_sha256': Value('string'), 'expected_answer_points': List(Value('string')), 'expected_behavior': Value('string'), 'supported_answer_points': List(Value('string'))}
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
pair_id: string
case_id: string
honest_arm: string
fabricated_arm: string
honest_trace_artifact: string
fabricated_trace_artifact: string
honest_trace_sha256: string
fabricated_trace_sha256: string
mutation_scope: string
fabrication_category: string
unsupported_figure_count: int64
changed_attribute_count: int64
verification_status: string
pair_schema_version: string
trajectory_expectations: null
title: string
page_anchors_pdf: list<item: int64>
child 0, item: int64
supported_answer_points: list<item: string>
child 0, item: string
definition_sensitive: bool
expected_answer_points: list<item: string>
child 0, item: string
case_type: string
corpus_support: string
expected_outcome_class: string
question: string
corpus_id: string
expected_behavior: string
source_case_sha256: string
case_schema_version: string
to
{'case_id': Value('string'), 'title': Value('string'), 'question': Value('string'), 'case_type': Value('string'), 'expected_outcome_class': Value('string'), 'corpus_support': Value('string'), 'definition_sensitive': Value('bool'), 'page_anchors_pdf': List(Value('int64')), 'trajectory_expectations': Json(decode=True), 'corpus_id': Value('string'), 'case_schema_version': Value('string'), 'source_case_sha256': Value('string'), 'expected_answer_points': List(Value('string')), 'expected_behavior': Value('string'), 'supported_answer_points': List(Value('string'))}
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.
Plumloom RAG Evaluation Benchmark
This dataset packages the public tables from a corpus-boundary study of a document-grounded RAG agent over NIST SP 800-160 Vol. 1 Rev. 1. The research asks whether an enterprise team can determine that a document agent is operating from its designated corpus—and turn observed behavior into reliable release evidence.
The benchmark creates difficult corpus-boundary tests. The RAG produces the behavior. EvalEngine turns those executions into trace-level quality evidence, judge-replicated measurements, and multi-run reliability signals that support release decisions.
Files
| File | Rows or purpose |
|---|---|
cases.jsonl |
20 public, corpus-verified benchmark cases |
matched-pairs.jsonl |
Eight honest/fabricated matched-pair relationships |
research-trace-results.csv |
32 case-arm results across two judge protocols |
research-scenario-results.csv |
28 metric-level N16 Scenario aggregate rows |
scenario-config.json |
Shared, redacted N16 Scenario configuration |
checksums.sha256 |
SHA-256 checksums for the six other files |
case_id joins cases to trace results. pair_id joins the matched-pair table
to the honest and fabricated result rows. Scenario rows are filtered by
information_state, evaluation_role, and metric.
Methodology
The 20-case benchmark combines a corpus term-probe audit, a strategic behavior taxonomy, definition-shift tests, and deterministic pre-flight verification. The agent used BM25 retrieval and an iterative retrieve-assess-refine-answer loop with citations and deterministic grounding checks.
Eight source traces were paired with controlled fabricated twins. Within each pair, the retrieval trajectory was held fixed while unsupported numerical claims and their grounding attributes were added to the final answer. The same frozen artifacts were evaluated under an initial DeepSeek V4 Pro protocol and a second-judge Kimi K3 replication.
For N16, Scenario evaluation compared a retrieved-only information state with a glossary-focused information state. Repeated runs produced means, variance, coefficient of variation, convergence status, and 95% confidence intervals. Scenario evaluation supplied reference material directly; it did not execute the live RAG retrieval loop.
Headline findings
- The RAG produced the expected outcome class in 19 of 20 cases, including all five expected refusal behaviors.
- The fabricated arm scored below its matched honest arm in all eight pairs under both judges: 16 of 16 paired comparisons.
- Factuality and Relevance produced the largest mean paired gaps under both judge configurations.
- N16 trace evaluation exposed a retrieval/trajectory failure despite a passing deterministic grounding flag.
- The N16 Scenario conditions converged, and independent replicas produced closely aligned configuration-level means.
Intended use and limitations
The data supports research on document-agent evaluation, RAG regression tests, corpus faithfulness, matched-pair evaluator validation, judge replication, and multi-run reliability analysis. It does not establish universal model or judge rankings, performance on corpora other than this NIST publication, or evidence that the underlying RAG was improved.
The evidence covers one 195-page publication, 20 cases, eight matched pairs, two judge families, and one Scenario case. The two N16 information states also differ in relevant coverage and surrounding context, so their comparison is descriptive rather than causal.
Provenance and licensing
Plumloom-authored metadata and documentation, generated model outputs, and NIST-derived material have distinct provenance and may require distinct terms. No blanket license is asserted here. The NIST PDF, source traces, code, logs, internal reports, platform records, and prompt-injection artifacts are not included in this upload package.
Research attribution: Plumloom Research. Source corpus: NIST SP 800-160 Vol. 1 Rev. 1, Engineering Trustworthy Secure Systems (November 2022).
- Downloads last month
- 40