README.md

metadata

license: cc0-1.0
language:
  - en
  - zh
  - hi
  - es
  - fr
  - ar
  - bn
  - pt
  - ru
  - ur
  - id
  - de
  - ja
  - pcm
  - mr
  - te
  - ta
  - tr
  - vi
  - ko
  - fa
pretty_name: Unit Conversion (21 languages)
size_categories:
  - 10K<n<100K
task_categories:
  - question-answering
  - token-classification
tags:
  - units
  - unit-conversion
  - math
  - multilingual
  - synthetic
configs:
  - config_name: units
    data_files: units.csv
  - config_name: qa
    data_files:
      - split: train
        path: qa/train.csv
      - split: validation
        path: qa/validation.csv
      - split: test
        path: qa/test.csv
  - config_name: ner
    data_files:
      - split: train
        path: ner/train.jsonl
      - split: validation
        path: ner/validation.jsonl
      - split: test
        path: ner/test.jsonl

Unit Conversion (21 languages)

A three-part dataset about unit conversion:

1. units: a lookup table of 120 units with exact conversion factors to SI, offsets (temperature), names in 21 languages, aliases, and notes.
2. qa: 59,745 conversion questions in 21 languages, each with a numeric answer and a worked solution.
3. ner: 16,800 short texts in which quantities and units are annotated with character spans.

Every number is computed by code with exact rational arithmetic (fractions.Fraction) and rounded to 6 significant digits. No language model produced any answer. The generator is included (build_dataset.py, langdata_names.py, langdata_templates.py), so the whole dataset is reproducible (seed 20261105).

Loading

from datasets import load_dataset
units = load_dataset("YOUR_HF_USERNAME/unit-conversion", "units")["train"]
qa    = load_dataset("YOUR_HF_USERNAME/unit-conversion", "qa")
ner   = load_dataset("YOUR_HF_USERNAME/unit-conversion", "ner")

Languages

code	language	QA rows	NER rows	my confidence in the text
en	English	2,845	800	high
zh	Chinese (Mandarin)	2,845	800	high
hi	Hindi	2,845	800	high
es	Spanish	2,845	800	high
fr	French	2,845	800	high
ar	Arabic	2,845	800	high
bn	Bengali	2,845	800	lower
pt	Portuguese	2,845	800	high
ru	Russian	2,845	800	high
ur	Urdu	2,845	800	lower
id	Indonesian	2,845	800	high
de	German	2,845	800	high
ja	Japanese	2,845	800	high
pcm	Nigerian Pidgin	2,845	800	lower
mr	Marathi	2,845	800	lower
te	Telugu	2,845	800	lower
ta	Tamil	2,845	800	lower
tr	Turkish	2,845	800	high
vi	Vietnamese	2,845	800	high
ko	Korean	2,845	800	high
fa	Persian	2,845	800	high

One QA/NER row exists per language, not as translations of each other: values and templates are sampled independently, but the split is grouped by unit pair, so no pair is shared across splits in any language.

1. units.csv

SI value = value × to_si_factor + to_si_offset. Conversion between units A and B: (v × factor_A + offset_A − offset_B) / factor_B.

column	meaning
unit_id	stable identifier (used by the other parts)
dimension	length, mass, time, temperature, temperature_difference, area, volume, speed, pressure, energy, power, force, data
name_en, name_en_plural, name_<lang>, symbol	names (one column per language) and symbol
to_si_factor, to_si_factor_exact, to_si_offset	factor as float, as an exact fraction, and the offset
si_unit	the reference unit of the dimension (B for data)
system	SI, metric, imperial, imperial_us, us_customary, traditional, ...
aliases	`
trap_group	common confusion group, if any (see below)
reliability	exact, conventional, or traditional_variable
source, notes	provenance and remarks

2. qa/

column	meaning
id, lang (ISO code, 21 values), task	identifiers
question	the question text
answer_text, answer_value, answer_unit	answer as text, as a float (6 significant digits), and the target unit_id
solution	short worked solution in the question language
units_involved, dimension	units used and the dimension of the answer
trap_tags	traps present in the question (see below)
digits	latin or native (Persian, Arabic-Indic, Bengali or Devanagari digits)

Tasks and counts:

task	count
convert	45381
fuel_economy	504
word_problem_distance	4620
word_problem_energy	4620
word_problem_volume	4620

Splits: train 47,684, validation 6,118, test 5,943. Splits are grouped by unit pair (for example mi->km): a pair that appears in the test set never appears in train, in any language. For word problems the split is random. So the test set measures generalisation to unseen pairs, not memorisation.

3. ner/

Each row has text and entities, a list of {start, end, text, value, unit_id, dimension, si_value}. Offsets are Python character offsets. An entity spans the quantity and the unit together (for example ۵ کیلومتر). Splits: train 13,418, validation 1,689, test 1,693.

Traps covered (trap_tags)

• us_vs_uk_volume: US vs UK gallon, pint, fluid ounce.
• abs_vs_delta_temp: a temperature (20 °C → 68 °F) vs a temperature difference (20 °C → 36 °F).
• ton_variants: metric tonne, US short ton, UK long ton.
• binary_vs_decimal: GB vs GiB, MB vs MiB, ...
• calorie_variants: calorie vs kilocalorie.
• foot_variants, hp_variants, mmhg_vs_torr, oz_variants: international vs survey foot, mechanical vs metric horsepower, mmHg vs torr, avoirdupois vs troy ounce.
• regional_traditional: traditional units whose value changed by region and era.
• reciprocal_scale: fuel economy (mpg vs L/100 km), where the relation is not linear.
• multi_step: word problems that need several unit steps.

Limitations (please read)

• Traditional units are conventions. Values such as the mesqal (4.608 g), sir (75 g), man (Tabriz, 2.97 kg), zar (1.04 m), farsang (6 km) and jarib (1000 m²) are common modern values; historical values differed by city and era. They are marked traditional_variable.
• Text is templated and synthetic. Questions come from a small set of templates per language, so linguistic variety is lower than in real user text. All names and templates were written with the help of an AI assistant and need native-speaker review. Confidence is highest for the large languages and lower for Bengali, Urdu, Marathi, Telugu, Tamil and Nigerian Pidgin (marked lower above).
• Agreement and case. In Spanish, French, Portuguese, German and Russian the questions write units as symbols after numbers (5 km) to avoid plural and case errors; the units table still has full names. Arabic phrasing is simplified.
• Traditional Persian units outside Persian and Arabic are written in Latin transliteration (mesqal, sir, man, zar, farsang, jarib).
• Digits. Native digits are used only for Persian, Arabic, Urdu, Bengali, Hindi and Marathi (digits column); Tamil and Telugu use Latin digits.
• Some English units that no language has a short everyday name for (e.g. US survey foot) use descriptive names; compound names (square, cubic, per) are built by rule and can sound stiff.
• Fuel economy is in qa only (not in units) because it is not a linear scale.
• Answers are rounded to 6 significant digits; compare with a tolerance.
• Values with a conventional definition (mmHg, cal, BTU, Julian year) follow the cited convention.

License

CC0 1.0 (public domain dedication). Sources of constants: BIPM SI Brochure, NIST SP 811 and Handbook 44, IEC 80000-13.

https://toolshed.click/

license: cc0-1.0