Datasets:
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
anchor: string
positive: string
lang: string
negative: string
tier: string
eval_anchor_collisions_dropped: int64
tier_counts: struct<tier1: int64, fiqa: int64>
child 0, tier1: int64
child 1, fiqa: int64
base_encoder: string
matryoshka_dims: list<item: int64>
child 0, item: int64
lang_counts: struct<en: int64, en_x: int64, hi: int64, hi_x: int64, hing: int64, hing_x: int64>
child 0, en: int64
child 1, en_x: int64
child 2, hi: int64
child 3, hi_x: int64
child 4, hing: int64
child 5, hing_x: int64
num_en_variants_per_section: int64
num_sections: int64
report: string
seed: int64
num_hinglish_variants_per_section: int64
num_rows: int64
best_mean_recall@5: double
fiqa_cap: int64
ft_encoder: string
to
{'seed': Value('int64'), 'num_rows': Value('int64'), 'num_sections': Value('int64'), 'num_en_variants_per_section': Value('int64'), 'num_hinglish_variants_per_section': Value('int64'), 'lang_counts': {'en': Value('int64'), 'en_x': Value('int64'), 'hi': Value('int64'), 'hi_x': Value('int64'), 'hing': Value('int64'), 'hing_x': Value('int64')}, 'tier_counts': {'tier1': Value('int64'), 'fiqa': Value('int64')}, 'eval_anchor_collisions_dropped': Value('int64'), 'fiqa_cap': Value('int64'), 'ft_encoder': Value('string'), 'base_encoder': Value('string'), 'matryoshka_dims': List(Value('int64')), 'best_mean_recall@5': Value('float64'), 'report': 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
anchor: string
positive: string
lang: string
negative: string
tier: string
eval_anchor_collisions_dropped: int64
tier_counts: struct<tier1: int64, fiqa: int64>
child 0, tier1: int64
child 1, fiqa: int64
base_encoder: string
matryoshka_dims: list<item: int64>
child 0, item: int64
lang_counts: struct<en: int64, en_x: int64, hi: int64, hi_x: int64, hing: int64, hing_x: int64>
child 0, en: int64
child 1, en_x: int64
child 2, hi: int64
child 3, hi_x: int64
child 4, hing: int64
child 5, hing_x: int64
num_en_variants_per_section: int64
num_sections: int64
report: string
seed: int64
num_hinglish_variants_per_section: int64
num_rows: int64
best_mean_recall@5: double
fiqa_cap: int64
ft_encoder: string
to
{'seed': Value('int64'), 'num_rows': Value('int64'), 'num_sections': Value('int64'), 'num_en_variants_per_section': Value('int64'), 'num_hinglish_variants_per_section': Value('int64'), 'lang_counts': {'en': Value('int64'), 'en_x': Value('int64'), 'hi': Value('int64'), 'hi_x': Value('int64'), 'hing': Value('int64'), 'hing_x': Value('int64')}, 'tier_counts': {'tier1': Value('int64'), 'fiqa': Value('int64')}, 'eval_anchor_collisions_dropped': Value('int64'), 'fiqa_cap': Value('int64'), 'ft_encoder': Value('string'), 'base_encoder': Value('string'), 'matryoshka_dims': List(Value('int64')), 'best_mean_recall@5': Value('float64'), 'report': 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.
TaxRAG Fine-Tuning Pairs
Curated contrastive-retrieval training data for
vivekkopthsd/embeddinggemma-300m-taxrag-ft.
29,440 (anchor, positive, negative) rows — the exact schema consumed by
fine_tune_encoder.py. Deterministic (seed 7): every row is reproducible from
build_ft_pairs.py.
Files
| File | Contents |
|---|---|
ft_pairs.json.gz |
29,440 rows: {anchor, positive, negative, lang, tier} |
ft_meta.json |
counts, seed, generation config |
queries_en.json / queries_hi.json / queries_hinglish.json |
the 180 held-out eval queries (never in the pairs) |
Row schema
{"anchor": "How do I claim the deduction for life insurance premium?",
"positive": "Section 123: Deduction for life insurance premium ...",
"negative": "Section 122: Deductions to be made in computing total income ...",
"lang": "en_x", "tier": "tier1"}
- Positive — gold passage: EN or HI chunk of the anchor's section.
- Negative — BM25 top-1 non-gold passage from the same bilingual index (hard negatives against confusable sections).
lang—en/en_x(EN anchor → EN/HI passage),hi/hi_x(HI anchor → HI/EN passage),hing/hing_x(Latin-transliterated Hinglish anchor → EN/HI passage).tier—tier1(bilingual statute) orfiqa(general finance).
Distribution
- Tier 1 (19,440) — Income-Tax Act, 2025: 477 sections, 1,237 EN + 1,237 Hindi passages; 3 EN + 3 HI + 2 Hinglish anchors per section, each paired with both language variants (cross-lingual alignment).
- Tier 2 (10,000) — sampled from
BeIR/fiqatrain+validation qrels, with seeded random non-gold negatives.
License
Mixed: Tier 1 derives from an MIT-licensed statute-text dataset plus machine
translations; Tier 2 from BeIR/fiqa (CC-BY-SA-4.0). Published as
CC-BY-SA-4.0 (the more restrictive of the two).
- Downloads last month
- 15