Datasets:

ahmad-awad
/

arabic-normalization-retrieval

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
qid: string
query: string
tags: list<item: string>
  child 0, item: string
corpus_selection: struct<method: string, requested_distractors: null, scanned_passages: int64, included_passages: int6 (... 24 chars omitted)
  child 0, method: string
  child 1, requested_distractors: null
  child 2, scanned_passages: int64
  child 3, included_passages: int64
  child 4, label_informed: bool
scope: string
license_notice: string
split: string
query_selection: struct<method: string, seed: int64, requested_limit: null, available_queries: int64, selected_querie (... 44 chars omitted)
  child 0, method: string
  child 1, seed: int64
  child 2, requested_limit: null
  child 3, available_queries: int64
  child 4, selected_queries: int64
  child 5, selected_qids: list<item: string>
      child 0, item: string
source: string
dataset_revision: string
prepared_hashes: struct<corpus.jsonl: string, queries.jsonl: string, qrels.tsv: string>
  child 0, corpus.jsonl: string
  child 1, queries.jsonl: string
  child 2, qrels.tsv: string
sources: struct<topics: string, qrels: string, corpus: list<item: string>>
  child 0, topics: string
  child 1, qrels: string
  child 2, corpus: list<item: string>
      child 0, item: string
corpus_scope: string
corpus_revision: string
language: string
to
{'source': Value('string'), 'language': Value('string'), 'split': Value('string'), 'scope': Value('string'), 'corpus_scope': Value('string'), 'dataset_revision': Value('string'), 'corpus_revision': Value('string'), 'query_selection': {'method': Value('string'), 'seed': Value('int64'), 'requested_limit': Value('null'), 'available_queries': Value('int64'), 'selected_queries': Value('int64'), 'selected_qids': List(Value('string'))}, 'corpus_selection': {'method': Value('string'), 'requested_distractors': Value('null'), 'scanned_passages': Value('int64'), 'included_passages': Value('int64'), 'label_informed': Value('bool')}, 'sources': {'topics': Value('string'), 'qrels': Value('string'), 'corpus': List(Value('string'))}, 'license_notice': Value('string'), 'prepared_hashes': {'corpus.jsonl': Value('string'), 'queries.jsonl': Value('string'), 'qrels.tsv': 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 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
              qid: string
              query: string
              tags: list<item: string>
                child 0, item: string
              corpus_selection: struct<method: string, requested_distractors: null, scanned_passages: int64, included_passages: int6 (... 24 chars omitted)
                child 0, method: string
                child 1, requested_distractors: null
                child 2, scanned_passages: int64
                child 3, included_passages: int64
                child 4, label_informed: bool
              scope: string
              license_notice: string
              split: string
              query_selection: struct<method: string, seed: int64, requested_limit: null, available_queries: int64, selected_querie (... 44 chars omitted)
                child 0, method: string
                child 1, seed: int64
                child 2, requested_limit: null
                child 3, available_queries: int64
                child 4, selected_queries: int64
                child 5, selected_qids: list<item: string>
                    child 0, item: string
              source: string
              dataset_revision: string
              prepared_hashes: struct<corpus.jsonl: string, queries.jsonl: string, qrels.tsv: string>
                child 0, corpus.jsonl: string
                child 1, queries.jsonl: string
                child 2, qrels.tsv: string
              sources: struct<topics: string, qrels: string, corpus: list<item: string>>
                child 0, topics: string
                child 1, qrels: string
                child 2, corpus: list<item: string>
                    child 0, item: string
              corpus_scope: string
              corpus_revision: string
              language: string
              to
              {'source': Value('string'), 'language': Value('string'), 'split': Value('string'), 'scope': Value('string'), 'corpus_scope': Value('string'), 'dataset_revision': Value('string'), 'corpus_revision': Value('string'), 'query_selection': {'method': Value('string'), 'seed': Value('int64'), 'requested_limit': Value('null'), 'available_queries': Value('int64'), 'selected_queries': Value('int64'), 'selected_qids': List(Value('string'))}, 'corpus_selection': {'method': Value('string'), 'requested_distractors': Value('null'), 'scanned_passages': Value('int64'), 'included_passages': Value('int64'), 'label_informed': Value('bool')}, 'sources': {'topics': Value('string'), 'qrels': Value('string'), 'corpus': List(Value('string'))}, 'license_notice': Value('string'), 'prepared_hashes': {'corpus.jsonl': Value('string'), 'queries.jsonl': Value('string'), 'qrels.tsv': Value('string')}}
              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.

Arabic Normalization Retrieval Experiment

This dataset contains the saved artifacts from a controlled Arabic text-normalization retrieval experiment.

The experiment used the complete MIRACL Arabic corpus and all 2,896 queries in the public development set. It compared raw text with conservative external normalization for BM25 and facebook/mcontriever-msmarco.

This is an evaluation on a visible public development set. It is not a private blind test.

Main result

Retriever	Raw nDCG@10	Conservative nDCG@10	Difference	Paired 95% interval
BM25	0.345234	0.360498	+0.015264	[0.010742, 0.019882]
Dense	0.364622	0.360294	-0.004328	[-0.009548, 0.000777]

Conservative normalization improved BM25 in this experiment. It did not show a reliable improvement for the dense retriever.

Data files

data/miracl-ar-dev/queries.jsonl: the exact 2,896-query population.
data/miracl-ar-dev/qrels.tsv: the exact relevance judgments.
data/miracl-ar-dev/provenance.json: pinned source revisions and preparation hashes.
data/diagnostic/: a small fictional fixture used only for code checks.
results/metrics/: aggregate metrics, per-query metrics, and exploratory failure analysis.
results/trec/: four top-100 rankings and independent trec_eval verification.
results/run-reports/: run summaries and compressed original run reports.
results/SHA256SUMS: checksums for the main artifacts.

The four TREC rankings are BM25 raw, BM25 conservative, dense raw, and dense conservative.

What is not included

The repository does not include the 2,061,414-passage MIRACL corpus, model weights, embedding caches, or BM25 indexes. These large upstream and generated files are not required to inspect the saved results.

Normalization

The conservative condition:

Removes Arabic diacritics from U+064B through U+0652.
Removes tatweel U+0640.
Maps أ, إ, and آ to ا.

Raw means no external Arabic normalization. The dense model still uses its fixed internal tokenizer.

Reproducibility

The source code, protocol, commands, and tests are available in the GitHub repository.

Pinned revisions:

MIRACL topics and qrels: 5be20db9509754dadad47689368639fcec739c00.
MIRACL corpus: d921ec7e349ce0d28daf30b2da9da5ee698bef0d.
Dense model: facebook/mcontriever-msmarco at 9ff6abed2c2fdf32bbbd8b4e98fb10160e317375.

The primary paired bootstrap used 10,000 samples and seed 42. NIST trec_eval independently checked 40 aggregate values. All checked values agreed within 0.00005.

Limitations

MIRACL relevance judgments are incomplete. An unjudged passage receives zero credit in the saved metrics, but this does not prove that a human judged it irrelevant.

The failure analysis was completed after the outcomes were visible. It is exploratory.

License and attribution

This dataset repository uses Apache License 2.0, matching the official MIRACL dataset repository. The diagnostic fixture and original project documentation were created for this experiment. MIRACL-derived files retain their upstream attribution requirements.

MIRACL citation

@article{10.1162_tacl_a_00595,
  author = {Zhang, Xinyu and Thakur, Nandan and Ogundepo, Odunayo and Kamalloo, Ehsan and Alfonso-Hermelo, David and Li, Xiaoguang and Liu, Qun and Rezagholizadeh, Mehdi and Lin, Jimmy},
  title = {MIRACL: A Multilingual Retrieval Dataset Covering 18 Diverse Languages},
  journal = {Transactions of the Association for Computational Linguistics},
  volume = {11},
  pages = {1114--1131},
  year = {2023},
  doi = {10.1162/tacl_a_00595}
}

Downloads last month: 69

Total file size:

76.4 MB