--- license: mit language: - en pretty_name: Calxo Unit Conversion Reference size_categories: - n<1K tags: - mathematics - unit-conversion - synthetic - evaluation configs: - config_name: default data_files: - split: validation path: conversion-reference.jsonl --- # Calxo Unit Conversion Reference 612 synthetic reference cases for numeric length and speed conversion. These cases support small regression checks of calculation software or numeric answers. They are not evidence of broad model reasoning ability, a comprehensive benchmark, or measured human activity. ## Task and format Each row includes an instruction, input magnitude, quantity, source and target unit, exact rational expected answer, floating-point approximation, tolerances and a definition source URL. Expected rational numerators and denominators are strings so integer precision survives common JSON readers. `case_id` is the stable key for predictions. There are 432 length cases and 180 speed cases, with inputs 0, 0.1, 0.5, 1, 22 and 100. Units and generation method are documented in the [source dataset](https://github.com/tigerman83/calxo-calculator-resources/tree/main/datasets/conversion-reference). This distribution adds task instructions, JSONL loading and a strict numeric evaluation recipe to that same reference collection. It is not a new independent source of ground truth. The single `validation` split is a reference partition. No training or held-out generalization split is claimed. Closely related rows share unit definitions, so a random split would leak those relationships. ## Load locally ```python import json from pathlib import Path rows = [json.loads(line) for line in Path("conversion-reference.jsonl").read_text(encoding="utf-8").splitlines()] assert len(rows) == 612 print(rows[0]["instruction"]) ``` Install the Hugging Face `datasets` package, then load the published validation split: ```python from datasets import load_dataset cases = load_dataset("rakeshvetal/calxo-unit-conversion-reference", split="validation") assert len(cases) == 612 ``` The dataset viewer is configured explicitly in the metadata above. ## Evaluate numeric predictions Write JSONL with one record per case: `{"case_id": "length-m-km-1-1", "prediction": 0.001}`. Predictions must be finite JSON numbers, with no missing, duplicate, extra or unknown case IDs. Run: ```sh python evaluate.py predictions.jsonl ``` The evaluator compares each prediction with `expected_numerator / expected_denominator` using `abs(actual - expected) <= max(absolute_tolerance, relative_tolerance * abs(expected))`. It reports passing cases and coverage. It intentionally does not extract numbers from prose; do any extraction separately and document that method. A perfect score on this small deterministic set does not establish general mathematical capability. ## Provenance and limitations Source: Calxo AI's independently authored reference collection, version 1.0.0. Expected answers were derived using rational arithmetic independently of the JavaScript implementation being checked. Definition sources include [NIST conversion factors](https://www.nist.gov/pml/special-publication-811/nist-guide-si-appendix-b-conversion-factors) and [NOAA nautical miles and knots](https://oceanservice.noaa.gov/facts/nautical-mile-knot.html). Those organizations do not endorse these examples. The cases exclude invalid inputs, extreme magnitudes, temperatures, historical miles, survey feet and condition-dependent Mach conversions. Tolerances apply to these example magnitudes; they are not measurement uncertainties. Inputs are non-negative magnitudes. The collection contains no user data or personal information. For wider calculation context, see the [Calxo Formula Library](https://calxo.ai/formula-library). An [executed Kaggle notebook](https://www.kaggle.com/code/rakeshvetal83/validate-calxo-unit-conversions) demonstrates a separate Python implementation and floating-point comparisons. ## License MIT applies to this authored distribution and evaluation script. The wider Formula Library archive has separate reuse terms. The Kaggle-hosted notebook copy is Apache 2.0. Preserve those distinctions when redistributing resources. Corrections can be reported in the public source repository.