---
license: cc-by-sa-4.0
language:
- is
- en
pretty_name: MGSM-IS (Icelandic MGSM)
task_categories:
- text-generation
- question-answering
tags:
- math
- mgsm
- gsm8k
- icelandic
- machine-translated
size_categories:
- n<1K
configs:
- config_name: default
  data_files:
  - split: test
    path: test.jsonl
---

# MGSM-IS: the MGSM math benchmark in Icelandic

> **Not manually verified yet.** Every question is a machine translation, chosen or corrected by an LLM adjudicator. No native speaker has reviewed the set. Expect some unidiomatic phrasing and possibly a few meaning errors. Report problems in the Community tab.

This is an Icelandic version of the 250 test questions of [MGSM](https://huggingface.co/datasets/juletxara/mgsm) (Shi et al., 2022). MGSM's questions are the first 250 questions of the [GSM8K](https://huggingface.co/datasets/openai/gsm8k) test set, so each item also records its GSM8K test index. The answers are the original numeric answers.

## How it was made (26 September 2026)

1. **Two independent translations** of each English question, at temperature 0.2 with the same instructions: keep every number, unit and name, write natural Icelandic, and render `$2` as "2 dollara". One came from Claude Opus 5.5 (`translation_opus_5_5`), the other from Gemini 3.8 Flash (`translation_gemini_3_8_flash`).
2. **Adjudication.** Claude Opus 5.5 compared both translations against the English source and chose A (Opus), B (Gemini) or its own corrected version. On these 250 items it chose A 164 times, B 72 times and edited 14. The chosen text is in `question`, and the verdict is in `adjudication`.
3. **Number check.** Every number in the English question must appear as digits in the translation. The flags that remain are all false positives: small numbers written as words (*tveir*, *fimm*), and US amounts such as `$.25` written *0,25 dollara*.

US units and currencies are kept, not localised, so scores stay comparable with English MGSM.

**Known limitations:**
- The adjudicator judged its own translation in position A, so self-preference and position bias are both possible. Neither has been measured.
- There has been no human review (see the warning above).

## Files and fields

- `test.jsonl`: `id` (0–249, the MGSM order), `question` (Icelandic), `answer_number`, `question_en` (MGSM English), `gsm8k_test_index`, `adjudication` (`A`, `B` or `edited`), `translation_opus_5_5`, `translation_gemini_3_8_flash`.
- `mgsm_is.tsv`: `question<TAB>answer_number`, the same layout as the original MGSM TSV files.

A suggested prompt that ends in a parseable answer: *"Leystu dæmið skref fyrir skref og endaðu svo á línunni 'Svar: N', þar sem N er lokatalan."*

## Licence and citation

The questions derive from MGSM (CC BY-SA 4.0), which derives from GSM8K (MIT), so this set is released under CC BY-SA 4.0. Please cite the originals:

```bibtex
@article{shi2022language,
  title={Language Models are Multilingual Chain-of-Thought Reasoners},
  author={Shi, Freda and Suzgun, Mirac and Freitag, Markus and Wang, Xuezhi and Srivats, Suraj and Vosoughi, Soroush and Chung, Hyung Won and Tay, Yi and Ruder, Sebastian and Zhou, Denny and Das, Dipanjan and Wei, Jason},
  journal={arXiv preprint arXiv:2210.03057},
  year={2022}
}
@article{cobbe2021gsm8k,
  title={Training Verifiers to Solve Math Word Problems},
  author={Cobbe, Karl and Kosaraju, Vineet and Bavarian, Mohammad and Chen, Mark and Jun, Heewoo and Kaiser, Lukasz and Plappert, Matthias and Tworek, Jerry and Hilton, Jacob and Nakano, Reiichiro and Hesse, Christopher and Schulman, John},
  journal={arXiv preprint arXiv:2110.14168},
  year={2021}
}
```