corruptedzayn/unit-conversion-reasoning

Unit Conversion Reasoning

A small, fully synthetic dataset of unit-conversion word problems with step-by-step worked solutions, covering length, weight, volume, temperature, and time.

Dataset Summary

Each example presents a unit-conversion problem (a direct conversion instruction, a chained multi-step conversion, or a short word problem) along with a step-by-step "reasoning" trace and a final "answer" string. The dataset was generated programmatically from a fixed table of conversion factors (and the standard Celsius/Fahrenheit/Kelvin formulas for temperature) — there is no scraped or human-written content.

• Total examples: 1,500 (deduplicated)
• Train / test split: 90% / 10% (1,350 / 150)
• Categories: length, weight, volume, temperature, time
• Difficulties: easy (single-step), medium (chained, two-step), hard (word-problem phrasing)

Schema

Field	Type	Description
question	string	The conversion problem, either a direct instruction or a word problem.
answer	string	The final numeric answer with its unit, e.g. "29.199475 feet".
reasoning	string	A step-by-step worked solution explaining the conversion.
category	string	One of length, weight, volume, temperature, time.
difficulty	string	One of easy, medium, hard.

Example row

{
  "question": "Convert 8.9 meters to feet.",
  "answer": "29.199475 feet",
  "reasoning": "Step 1: 1 meter = 1 meter, and 1 foot = 0.3048 of the same base unit, so 1 meter = 3.28084 feet.\nStep 2: 8.9 meters x 3.28084 = 29.199475 feet.",
  "category": "length",
  "difficulty": "easy"
}

Generation Method

The dataset is produced by generate_dataset.py, using fixed, standard conversion factor tables for length, weight, volume, and time (each expressed relative to a base unit — meter, gram, liter, second), and the standard linear/affine formulas for Celsius/Fahrenheit/Kelvin temperature conversion.

For each example, the script:

1. Picks a category and a difficulty level.
2. Easy — generates a single direct conversion between two random units in that category.
3. Medium — generates a chained conversion through an intermediate unit (e.g. inches → meters → kilometers), requiring two reasoning steps.
4. Hard — wraps a single or two-step conversion in a short word-problem template (e.g. "A recipe calls for 2.5 liters of milk, how many cups is that?").
5. Computes the exact numeric result from the conversion factors directly (not from the displayed reasoning text), formats the question/answer/reasoning strings, and records the category and difficulty.

Problems are deduplicated by exact question text. After deduplication, every row is passed through an independent checker (validate_row in the generation script) that re-parses the numeric value out of the answer string and confirms it matches the actual computed conversion to within a small numeric tolerance; any row that fails this check is dropped before the dataset is saved.

The dataset is purely synthetic and rule-based: no scraping, no external text sources, and no human annotation were used.

Intended Use

This dataset is intended for:

• Evaluating LLMs' quantitative reasoning and unit-conversion accuracy.
• Fine-tuning or instruction-tuning models to produce step-by-step arithmetic/conversion reasoning.
• Benchmarking chain-of-thought consistency (does the reasoning trace actually support the final answer?).

It is not intended as a general-purpose math benchmark — it covers a narrow domain (unit conversion) and uses a fixed set of templates and units, so models can in principle overfit to its surface patterns.

Limitations

• Templates and unit lists are fixed and limited, so question phrasing has noticeable repetition across the dataset.
• Numeric answers are rounded for display (typically to 6 decimal places or fewer); exact-string-match evaluation should allow for reasonable numeric tolerance rather than requiring exact text equality.
• Only "everyday" units are covered (no scientific/imperial obscure units like furlongs, slugs, etc.).

License

This dataset is released under CC0 1.0 (Public Domain Dedication). It is entirely synthetic, generated from arithmetic conversion formulas, and contains no personal, copyrighted, or scraped data.