Datasets:

shivam909067
/

indic-roman-gap

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
name: string
version: string
what: string
sources: struct<sarvamai/mmlu-indic: struct<licence: string, redistributed: bool>, sarvamai/gsm8k-indic: stru (... 55 chars omitted)
  child 0, sarvamai/mmlu-indic: struct<licence: string, redistributed: bool>
      child 0, licence: string
      child 1, redistributed: bool
  child 1, sarvamai/gsm8k-indic: struct<licence: string, redistributed: bool, note: string>
      child 0, licence: string
      child 1, redistributed: bool
      child 2, note: string
seed: int64
per_dir: struct<mmlu-indic: int64, gsm8k-indic: int64>
  child 0, mmlu-indic: int64
  child 1, gsm8k-indic: int64
languages: list<item: string>
  child 0, item: string
scripts: string
suites: struct<mmlu-indic: struct<n: int64, item_ids: list<item: string>>, gsm8k-indic: struct<n: int64, ite (... 27 chars omitted)
  child 0, mmlu-indic: struct<n: int64, item_ids: list<item: string>>
      child 0, n: int64
      child 1, item_ids: list<item: string>
          child 0, item: string
  child 1, gsm8k-indic: struct<n: int64, item_ids: list<item: string>>
      child 0, n: int64
      child 1, item_ids: list<item: string>
          child 0, item: string
gsm8k-indic: struct<model: string, overall: double, n: int64, by_lang: struct<bn: double, bn_roman: double, en: d (... 157 chars omitted)
  child 0, model: string
  child 1, overall: double
  child 2, n: int64
  child 3, by_lang: struct<bn: double, bn_roman: double, en: double, gu: double, gu_roman: double, hi: double, hi_roman: (... 98 chars omitted)
      child 0, bn: double
      child 1, bn_roman: double
      child 2, en: double
      child 3, gu: double
      child 4, gu_roman: double
      child 5, hi: double
      child 6, hi_roman: double
      child 7, mr: double
      child 8, mr_roman: double
      child 9, ta: double
      child 10, ta_roman: double
      child 11, te: double
      child 12, te_roman: double
mmlu-indic: struct<model: string, overall: double, n: int64, by_lang: struct<bn: double, bn_roman: double, en: d (... 157 chars omitted)
  child 0, model: string
  child 1, overall: double
  child 2, n: int64
  child 3, by_lang: struct<bn: double, bn_roman: double, en: double, gu: double, gu_roman: double, hi: double, hi_roman: (... 98 chars omitted)
      child 0, bn: double
      child 1, bn_roman: double
      child 2, en: double
      child 3, gu: double
      child 4, gu_roman: double
      child 5, hi: double
      child 6, hi_roman: double
      child 7, mr: double
      child 8, mr_roman: double
      child 9, ta: double
      child 10, ta_roman: double
      child 11, te: double
      child 12, te_roman: double
to
{'mmlu-indic': {'model': Value('string'), 'overall': Value('float64'), 'n': Value('int64'), 'by_lang': {'bn': Value('float64'), 'bn_roman': Value('float64'), 'en': Value('float64'), 'gu': Value('float64'), 'gu_roman': Value('float64'), 'hi': Value('float64'), 'hi_roman': Value('float64'), 'mr': Value('float64'), 'mr_roman': Value('float64'), 'ta': Value('float64'), 'ta_roman': Value('float64'), 'te': Value('float64'), 'te_roman': Value('float64')}}, 'gsm8k-indic': {'model': Value('string'), 'overall': Value('float64'), 'n': Value('int64'), 'by_lang': {'bn': Value('float64'), 'bn_roman': Value('float64'), 'en': Value('float64'), 'gu': Value('float64'), 'gu_roman': Value('float64'), 'hi': Value('float64'), 'hi_roman': Value('float64'), 'mr': Value('float64'), 'mr_roman': Value('float64'), 'ta': Value('float64'), 'ta_roman': Value('float64'), 'te': Value('float64'), 'te_roman': 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
              name: string
              version: string
              what: string
              sources: struct<sarvamai/mmlu-indic: struct<licence: string, redistributed: bool>, sarvamai/gsm8k-indic: stru (... 55 chars omitted)
                child 0, sarvamai/mmlu-indic: struct<licence: string, redistributed: bool>
                    child 0, licence: string
                    child 1, redistributed: bool
                child 1, sarvamai/gsm8k-indic: struct<licence: string, redistributed: bool, note: string>
                    child 0, licence: string
                    child 1, redistributed: bool
                    child 2, note: string
              seed: int64
              per_dir: struct<mmlu-indic: int64, gsm8k-indic: int64>
                child 0, mmlu-indic: int64
                child 1, gsm8k-indic: int64
              languages: list<item: string>
                child 0, item: string
              scripts: string
              suites: struct<mmlu-indic: struct<n: int64, item_ids: list<item: string>>, gsm8k-indic: struct<n: int64, ite (... 27 chars omitted)
                child 0, mmlu-indic: struct<n: int64, item_ids: list<item: string>>
                    child 0, n: int64
                    child 1, item_ids: list<item: string>
                        child 0, item: string
                child 1, gsm8k-indic: struct<n: int64, item_ids: list<item: string>>
                    child 0, n: int64
                    child 1, item_ids: list<item: string>
                        child 0, item: string
              gsm8k-indic: struct<model: string, overall: double, n: int64, by_lang: struct<bn: double, bn_roman: double, en: d (... 157 chars omitted)
                child 0, model: string
                child 1, overall: double
                child 2, n: int64
                child 3, by_lang: struct<bn: double, bn_roman: double, en: double, gu: double, gu_roman: double, hi: double, hi_roman: (... 98 chars omitted)
                    child 0, bn: double
                    child 1, bn_roman: double
                    child 2, en: double
                    child 3, gu: double
                    child 4, gu_roman: double
                    child 5, hi: double
                    child 6, hi_roman: double
                    child 7, mr: double
                    child 8, mr_roman: double
                    child 9, ta: double
                    child 10, ta_roman: double
                    child 11, te: double
                    child 12, te_roman: double
              mmlu-indic: struct<model: string, overall: double, n: int64, by_lang: struct<bn: double, bn_roman: double, en: d (... 157 chars omitted)
                child 0, model: string
                child 1, overall: double
                child 2, n: int64
                child 3, by_lang: struct<bn: double, bn_roman: double, en: double, gu: double, gu_roman: double, hi: double, hi_roman: (... 98 chars omitted)
                    child 0, bn: double
                    child 1, bn_roman: double
                    child 2, en: double
                    child 3, gu: double
                    child 4, gu_roman: double
                    child 5, hi: double
                    child 6, hi_roman: double
                    child 7, mr: double
                    child 8, mr_roman: double
                    child 9, ta: double
                    child 10, ta_roman: double
                    child 11, te: double
                    child 12, te_roman: double
              to
              {'mmlu-indic': {'model': Value('string'), 'overall': Value('float64'), 'n': Value('int64'), 'by_lang': {'bn': Value('float64'), 'bn_roman': Value('float64'), 'en': Value('float64'), 'gu': Value('float64'), 'gu_roman': Value('float64'), 'hi': Value('float64'), 'hi_roman': Value('float64'), 'mr': Value('float64'), 'mr_roman': Value('float64'), 'ta': Value('float64'), 'ta_roman': Value('float64'), 'te': Value('float64'), 'te_roman': Value('float64')}}, 'gsm8k-indic': {'model': Value('string'), 'overall': Value('float64'), 'n': Value('int64'), 'by_lang': {'bn': Value('float64'), 'bn_roman': Value('float64'), 'en': Value('float64'), 'gu': Value('float64'), 'gu_roman': Value('float64'), 'hi': Value('float64'), 'hi_roman': Value('float64'), 'mr': Value('float64'), 'mr_roman': Value('float64'), 'ta': Value('float64'), 'ta_roman': Value('float64'), 'te': Value('float64'), 'te_roman': 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.

indic-roman-gap

How much worse is an Indic LLM when you type its language in Latin letters?

Hundreds of millions of people write Hindi, Gujarati, Tamil and Marathi in the Latin alphabet every day — on WhatsApp, Instagram, Telegram. This is a frozen, reproducible measurement of what that costs.

Measured on sarvam-m (24B), the strongest openly-licensed Indic model:

language	native	romanised	gap
Tamil	0.790	0.570	−22.0
Marathi	0.760	0.630	−13.0
Bengali	0.840	0.720	−12.0
Hindi	0.870	0.790	−8.0
Gujarati	0.790	0.710	−8.0
Telugu	0.790	0.710	−8.0

GSM8K-Indic, 1,300 items. Knowledge (MMLU-Indic, 2,600 items) shows the same direction: −5.7pp, averaged over the six per-language gaps.

Every language. Both suites. Same questions, only the script changes.

Why this is interesting

It is a format problem, not a knowledge problem. The model answers Gujarati maths correctly 79% of the time in Gujarati script and 71% of the time when the identical question arrives in Latin letters. The knowledge is there; the surface form costs it.

Romanised Indic is not transliteration. Look at a real item:

Ek vyakti television par volume star badalne ke liye
remote control ka upyog karta hai.

Phonetic spelling, no diacritics, English technical terms left in English. No transliteration scheme produces this, which is why it has to be learned and why models trained on native script alone do not get it for free.

Native-script Indic reasoning already beats English. sarvam-m scores 0.807 on Indic-native maths against 0.780 on English. These are not translation layers over an English model. The weakness is specifically the Latin-script surface.

What this repo contains

frozen_selection.json     item ids + seed for an exact, frozen sample
baseline_sarvam_m.json    our measured baseline, per language and script
build_frozen_suite.py     rebuilds the sample from the public sources
run_frozen.py             the scorer

No dataset content is redistributed. sarvamai/mmlu-indic is MIT, but sarvamai/gsm8k-indic states no licence on the Hub, so this publishes the selection — ids, seed, recipe — and you reconstruct it locally from the originals. Same items, byte for byte, no licence question.

Reproducing it

hf download sarvamai/mmlu-indic  --repo-type dataset --local-dir data/mmlu-indic
hf download sarvamai/gsm8k-indic --repo-type dataset --local-dir data/gsm8k-indic
python build_frozen_suite.py          # seed 42, 200/100 items per language dir
python run_frozen.py --model <path> --label yours --suite mmlu-indic

Two methodology notes that change the numbers

Multiple choice is scored by likelihood, not generation. Scoring by generation measures whether the model emitted a parseable letter, which is formatting compliance rather than knowledge.

Reasoning models must have thinking disabled for multiple choice. sarvam-m emits a chain of thought by default, so the first generated position is the first token of its reasoning, not its answer. Scoring there returned 0.300 — chance on a 4-way question. With enable_thinking=False the same 26 items scored 0.615. If you benchmark a reasoning model, check this before publishing anything.

What this is not

A leaderboard. One model, one measurement, run by people who intend to improve on it — treat the numbers as reproducible rather than as authority, and re-run them yourself.

Licence

Code MIT. Baseline numbers CC0. No third-party dataset content included.

Downloads last month: 16

Total file size:

141 kB