Datasets:

baogiadoan
/

dora-benchmark

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
seed_id: string
task: string
question: string
answer: string
reference: list<item: string>
  child 0, item: string
references: list<item: struct<doc_id: string, doc_name: string, doc_path: string, page_numbers: list<item: int64 (... 38 chars omitted)
  child 0, item: struct<doc_id: string, doc_name: string, doc_path: string, page_numbers: list<item: int64>, chunk_id (... 26 chars omitted)
      child 0, doc_id: string
      child 1, doc_name: string
      child 2, doc_path: string
      child 3, page_numbers: list<item: int64>
          child 0, item: int64
      child 4, chunk_id: string
      child 5, snippet: string
metadata: string
synthesis: struct<mode: string, configured_min_chunks: int64, configured_max_chunks: int64, before_count: int64 (... 250 chars omitted)
  child 0, mode: string
  child 1, configured_min_chunks: int64
  child 2, configured_max_chunks: int64
  child 3, before_count: int64
  child 4, fallback_reason: null
  child 5, added_chunk_ids: list<item: string>
      child 0, item: string
  child 6, trimmed_chunk_ids: list<item: null>
      child 0, item: null
  child 7, token_budget_exceeded: bool
  child 8, after_count: int64
  child 9, unique_chunk_ids: list<item: string>
      child 0, item: string
  child 10, page_span: list<item: int64>
      child 0, item: int64
  child 11, source_sections: list<item: null>
      child 0, item: null
generation_stage: string
batch_number: int64
batch_index: int64
batch_duration: double
generation_method: string
qualit
...
 rag_overall: double
  child 8, strategic_overall: double
  child 9, strategic_vocabulary: double
  child 10, strategic_future_focus: double
  child 11, strategic_systems_thinking: double
  child 12, strategic_operational_context: double
  child 13, strategic_classification: string
composite_quality_score: double
meets_gates: bool
instruction: string
answer_with_keypoints: string
output: string
reference_keypoints: list<item: null>
  child 0, item: null
grounding_context: string
messages: list<item: struct<role: string, content: string>>
  child 0, item: struct<role: string, content: string>
      child 0, role: string
      child 1, content: string
keypoint_metrics: struct<completeness: double, hallucination: double, irrelevance: double, relevant_ids: list<item: in (... 88 chars omitted)
  child 0, completeness: double
  child 1, hallucination: double
  child 2, irrelevance: double
  child 3, relevant_ids: list<item: int64>
      child 0, item: int64
  child 4, irrelevant_ids: list<item: int64>
      child 0, item: int64
  child 5, wrong_ids: list<item: null>
      child 0, item: null
  child 6, responses: string
original_seed_id: string
wrong_ids: list<item: null>
  child 0, item: null
keypoints: list<item: string>
  child 0, item: string
responses: string
attempt_duration: double
completeness: double
attempt_number: int64
irrelevant_ids: list<item: int64>
  child 0, item: int64
relevant_ids: list<item: int64>
  child 0, item: int64
hallucination: double
irrelevance: double
to
{'seed_id': Value('string'), 'task': Value('string'), 'question': Value('string'), 'answer': Value('string'), 'reference': List(Value('string')), 'references': List({'doc_id': Value('string'), 'doc_name': Value('string'), 'doc_path': Value('string'), 'page_numbers': List(Value('int64')), 'chunk_id': Value('string'), 'snippet': Value('string')}), 'metadata': Json(decode=True), 'generation_stage': Value('string'), 'batch_number': Value('int64'), 'batch_index': Value('int64'), 'batch_duration': Value('float64'), 'generation_method': Value('string'), 'quality_metrics': {'rag_document_entailment': Value('float64'), 'rag_contradiction_penalty': Value('float64'), 'rag_span_f1': Value('float64'), 'rag_context_recall': Value('float64'), 'rag_context_precision': Value('float64'), 'rag_answer_relevancy': Value('float64'), 'rag_numeric_consistency': Value('float64'), 'rag_overall': Value('float64'), 'strategic_overall': Value('float64'), 'strategic_vocabulary': Value('float64'), 'strategic_future_focus': Value('float64'), 'strategic_systems_thinking': Value('float64'), 'strategic_operational_context': Value('float64'), 'strategic_classification': Value('string')}, 'composite_quality_score': Value('float64'), 'meets_gates': Value('bool'), 'keypoints': List(Value('string')), 'keypoint_metrics': {'completeness': Value('float64'), 'hallucination': Value('float64'), 'irrelevance': Value('float64'), 'relevant_ids': List(Value('int64')), 'irrelevant_ids': List(Value('int64')), 'wrong_ids': List(Value('null')), 'responses': Value('string')}, 'completeness': Value('float64'), 'hallucination': Value('float64'), 'irrelevance': Value('float64'), 'relevant_ids': List(Value('int64')), 'irrelevant_ids': List(Value('int64')), 'wrong_ids': List(Value('null')), 'responses': Value('string'), 'original_seed_id': Value('string'), 'messages': List({'role': Value('string'), 'content': Value('string')}), 'attempt_number': Value('int64'), 'attempt_duration': 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
              seed_id: string
              task: string
              question: string
              answer: string
              reference: list<item: string>
                child 0, item: string
              references: list<item: struct<doc_id: string, doc_name: string, doc_path: string, page_numbers: list<item: int64 (... 38 chars omitted)
                child 0, item: struct<doc_id: string, doc_name: string, doc_path: string, page_numbers: list<item: int64>, chunk_id (... 26 chars omitted)
                    child 0, doc_id: string
                    child 1, doc_name: string
                    child 2, doc_path: string
                    child 3, page_numbers: list<item: int64>
                        child 0, item: int64
                    child 4, chunk_id: string
                    child 5, snippet: string
              metadata: string
              synthesis: struct<mode: string, configured_min_chunks: int64, configured_max_chunks: int64, before_count: int64 (... 250 chars omitted)
                child 0, mode: string
                child 1, configured_min_chunks: int64
                child 2, configured_max_chunks: int64
                child 3, before_count: int64
                child 4, fallback_reason: null
                child 5, added_chunk_ids: list<item: string>
                    child 0, item: string
                child 6, trimmed_chunk_ids: list<item: null>
                    child 0, item: null
                child 7, token_budget_exceeded: bool
                child 8, after_count: int64
                child 9, unique_chunk_ids: list<item: string>
                    child 0, item: string
                child 10, page_span: list<item: int64>
                    child 0, item: int64
                child 11, source_sections: list<item: null>
                    child 0, item: null
              generation_stage: string
              batch_number: int64
              batch_index: int64
              batch_duration: double
              generation_method: string
              qualit
              ...
               rag_overall: double
                child 8, strategic_overall: double
                child 9, strategic_vocabulary: double
                child 10, strategic_future_focus: double
                child 11, strategic_systems_thinking: double
                child 12, strategic_operational_context: double
                child 13, strategic_classification: string
              composite_quality_score: double
              meets_gates: bool
              instruction: string
              answer_with_keypoints: string
              output: string
              reference_keypoints: list<item: null>
                child 0, item: null
              grounding_context: string
              messages: list<item: struct<role: string, content: string>>
                child 0, item: struct<role: string, content: string>
                    child 0, role: string
                    child 1, content: string
              keypoint_metrics: struct<completeness: double, hallucination: double, irrelevance: double, relevant_ids: list<item: in (... 88 chars omitted)
                child 0, completeness: double
                child 1, hallucination: double
                child 2, irrelevance: double
                child 3, relevant_ids: list<item: int64>
                    child 0, item: int64
                child 4, irrelevant_ids: list<item: int64>
                    child 0, item: int64
                child 5, wrong_ids: list<item: null>
                    child 0, item: null
                child 6, responses: string
              original_seed_id: string
              wrong_ids: list<item: null>
                child 0, item: null
              keypoints: list<item: string>
                child 0, item: string
              responses: string
              attempt_duration: double
              completeness: double
              attempt_number: int64
              irrelevant_ids: list<item: int64>
                child 0, item: int64
              relevant_ids: list<item: int64>
                child 0, item: int64
              hallucination: double
              irrelevance: double
              to
              {'seed_id': Value('string'), 'task': Value('string'), 'question': Value('string'), 'answer': Value('string'), 'reference': List(Value('string')), 'references': List({'doc_id': Value('string'), 'doc_name': Value('string'), 'doc_path': Value('string'), 'page_numbers': List(Value('int64')), 'chunk_id': Value('string'), 'snippet': Value('string')}), 'metadata': Json(decode=True), 'generation_stage': Value('string'), 'batch_number': Value('int64'), 'batch_index': Value('int64'), 'batch_duration': Value('float64'), 'generation_method': Value('string'), 'quality_metrics': {'rag_document_entailment': Value('float64'), 'rag_contradiction_penalty': Value('float64'), 'rag_span_f1': Value('float64'), 'rag_context_recall': Value('float64'), 'rag_context_precision': Value('float64'), 'rag_answer_relevancy': Value('float64'), 'rag_numeric_consistency': Value('float64'), 'rag_overall': Value('float64'), 'strategic_overall': Value('float64'), 'strategic_vocabulary': Value('float64'), 'strategic_future_focus': Value('float64'), 'strategic_systems_thinking': Value('float64'), 'strategic_operational_context': Value('float64'), 'strategic_classification': Value('string')}, 'composite_quality_score': Value('float64'), 'meets_gates': Value('bool'), 'keypoints': List(Value('string')), 'keypoint_metrics': {'completeness': Value('float64'), 'hallucination': Value('float64'), 'irrelevance': Value('float64'), 'relevant_ids': List(Value('int64')), 'irrelevant_ids': List(Value('int64')), 'wrong_ids': List(Value('null')), 'responses': Value('string')}, 'completeness': Value('float64'), 'hallucination': Value('float64'), 'irrelevance': Value('float64'), 'relevant_ids': List(Value('int64')), 'irrelevant_ids': List(Value('int64')), 'wrong_ids': List(Value('null')), 'responses': Value('string'), 'original_seed_id': Value('string'), 'messages': List({'role': Value('string'), 'content': Value('string')}), 'attempt_number': Value('int64'), 'attempt_duration': 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.

DoRA Benchmark

Benchmark instances from DoRA (Domain-oriented RAG Assessment), a benchmark construction and evaluation framework for specialist-domain RAG, instantiated on 40 public Australian defence-related documents.

📄 Paper · 💻 Code · 🤖 LoRA adapter

Splits

Split	Instances	Generator	Seed corpus
`test`	1,259	GPT-4o	20 documents
`train`	5,052	Claude Sonnet	20 documents (disjoint from test)
`validation`	266	Claude Sonnet	same as train
`expert` / `test`	86	Human domain experts	—

The train and test splits use different LLM families over disjoint seed documents. This cross-generator, cross-corpus design means a model fine-tuned on the training split cannot win on the test split by imitating the test generator's style or by memorising test-corpus content.

Intent styles

Every instance carries one of five practitioner-aligned styles:

Style	Test instances	Character
FIND	421	Single extractive fact
GENERATE	309	Enumeration / list synthesis
SUMMARIZE	209	Overview across passages
EXPLAIN	165	Concept, definition, relationship
PROVIDE	155	Quantitative / measurable data

Fields

Field	Description
`seed_id`	Source document identifier
`task`	Intent style (one of the five above)
`question`	The generated question
`answer`	Reference answer
`reference` / `references`	Supporting evidence passages with document and chunk provenance
`keypoints`	Rubric keypoints used by the faithfulness judge
`composite_quality_score`, `quality_metrics`	Construction-time quality scores

Instances are auditable: each carries the evidence bundle it was generated from, so any answer can be traced back to its supporting passages.

Licensing — please read

The project's own derived fields are released under CC BY 4.0. That covers the questions, answers, style labels, keypoints, quality scores and provenance structure authored by this project.

It does not license the underlying source documents. The reference / references fields contain short extracts from Australian Government publications that remain under their publishers' terms. Of the 40 seed documents, only 5 carry an explicit open licence; the rest assert Commonwealth copyright without an open grant, and defence.gov.au permits reproduction "in unaltered form for personal and non-commercial use". The source PDFs themselves are not redistributed here — the code repository ships a pointer manifest (seed_documents.jsonl, included here for convenience) with per-document URLs, SHA-256 hashes and rights status so each document can be retrieved from its publisher.

If you redistribute derivatives of this dataset, credit the source publications appropriately and do not represent the extracted passages as CC BY 4.0 material.

Evaluating with this benchmark

The RAG-faithfulness metrics (Completeness / Hallucination / Irrelevance) are produced by the RAGEval keypoint rubric judge, which is an external dependency licensed CC BY-NC-SA 4.0 and is not part of this release. Reproducing those columns inherits its NonCommercial term. Answer-coverage metrics (Token Recall, ROUGE-L Recall, BERTScore Recall) have no such restriction. See the code repository for setup.

Citation

@misc{doan2026benchmarkconstructionevaluationframework,
      title={A Benchmark Construction and Evaluation Framework for Specialist Domains: Case Study on Defense-related Documents},
      author={Bao Gia Doan and Aditya Joshi and Pantelis Elinas and Aarya Bodhankar and Oscar Leslie and Tom Marchant and Flora Salim},
      year={2026},
      eprint={2604.17943},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2604.17943},
}

Downloads last month: 12

Number of rows:

Total file size:

147 MB

Paper for baogiadoan/dora-benchmark

A Benchmark Construction and Evaluation Framework for Specialist Domains: Case Study on Defense-related Documents

Paper • 2604.17943 • Published May 27