danish-foundation-models/multilingual-gsm-symbolic

Multilingual GSM-Symbolic

Version	v0.5.5
Released	2026-10-06
Languages	108 (19 validated)

Multilingual GSM-Symbolic is a benchmark for evaluating arithmetic reasoning in large language models across many languages. It extends Apple's GSM-Symbolic approach by providing symbolic templates that generate thousands of structurally equivalent but numerically distinct math problems. Templates and generation are handled by the multilingual-gsm-symbolic package.

The dataset lets you test whether a model genuinely understands a problem or merely pattern-matches on the specific numbers it saw during training. Comparing the original and synthetic splits directly measures this gap.

Loading the dataset

With 🤗 datasets

from datasets import load_dataset

# English synthetic test split, pinned to a version for reproducible results (see the version table at the top)
eng = load_dataset("danish-foundation-models/multilingual-gsm-symbolic", name="eng", split="test_synthetic", revision="v0.5.0")

# Original test split of a language that has not yet been validated
swh = load_dataset("danish-foundation-models/multilingual-gsm-symbolic", name="swh", split="test_original_unvalidated", revision="v0.5.0")

With inspect-ai

You can evaluate with inspect-ai using the eval.yaml in this repository:

inspect eval hf/danish-foundation-models/multilingual-gsm-symbolic/synthetic_eng --model openai/gpt-5.4-nano --reasoning-effort low

To pin a version, add it after the task id: hf/danish-foundation-models/multilingual-gsm-symbolic/[email protected].

Task ids are original_<language> and synthetic_<language> for validated languages (e.g. original_deu) and original_unvalidated_<language> and synthetic_unvalidated_<language> for the rest. Omit the task id to run all of them.

Each task prompts the model in the language of the problem, asking it to solve the problem step by step and put the final answer in \boxed{}. The instructions are part of the templates in the GitHub repository (instruction.toml in each language folder) and can be loaded with multilingual_gsm_symbolic.load_instruction. Most instructions are machine translated and are annotated as validated once checked by a native speaker.

Dataset structure

Each language is a separate config (subset) named by its three-letter language code (e.g. eng, dan). eng_metric is a variant of English using metric units.

Split	Description
test_original	The concrete GSM problems for that language (100 for validated languages)
test_synthetic	20 generated variants per template (2,000 problems for validated languages)
test_original_unvalidated, test_synthetic_unvalidated	The same splits for languages that have not yet been validated by a native speaker. Templates that could not be translated reliably are left out, so some languages have fewer than 100
train_original	English only: 191 concrete training problems from GSM8K
train_synthetic	English only: 20 generated variants per training template (3,820 problems). The training templates are machine-generated and not validated

Fields

Field	Type	Description
question	string	The math problem
answer	string	Step-by-step solution ending with #### <number>
target	string	The final numeric answer (extracted from answer)
language	string	Three-letter language code (e.g. eng, dan)
source_id	int	Problem index in the original GSM8K dataset

Answer format

Answers follow the GSM8K convention — reasoning steps followed by a final numeric answer:

At 3 miles/hour, it will take 42/3=14 hours for the fog to cover the city.
#### 14

Languages

The dataset contains 19 validated languages, where every template has been computationally validated and checked by native speakers:

Language	Computationally validated	Human validated	Error analysis
ara	✓	by a native speaker	✓
dan	✓	by three native speakers	✓
deu	✓	by two native speakers	✓
eng	✓	by a native speaker	✓
eng_metric	✓	by a native speaker
est	✓	by a native speaker	✓
fra	✓	by a native speaker	✓
hin	✓	by a native speaker	✓
isl	✓	by native speakers	✓
ita	✓	by two native speakers	✓
jpn	✓	by a native speaker	✓
mar	✓	by a native speaker	✓
nld	✓	by a native speaker	✓
pan	✓	by two native speakers	✓
rus	✓	by a native speaker	✓
swe	✓	by 5 native speakers	✓
ukr	✓	by a native speaker	✓
urd	✓	by a native speaker	✓
zho	✓	by a native speaker	✓

The remaining 89 languages are machine translated and only computationally validated. Their splits are suffixed with _unvalidated: afr, amh, arz, asm, azb, bam, bar, bel, ben, bho, bos, bul, cat, ceb, ces, ctg, dyu, ell, fin, fuv, gaz, gle, guj, hat, hau, hne, hrv, hun, ibo, ind, jav, kan, kaz, khm, kin, kmr, kor, ktu, lav, lin, lit, lug, mag, mai, mal, mlt, mos, mya, nno, nob, npi, nso, nya, ory, pbu, pcm, pes, pnb, pol, por, ron, run, sck, sin, skr, slk, slv, sna, snd, som, spa, sun, swh, tam, tel, tgl, tha, tsn, tur, uig, uzn, vie, vjk, wol, xho, yor, yue, zlm, zul.

The evaluation instructions (see eval.yaml) have been checked by native speakers for: ara, dan, deu, eng, eng_metric, est, fra, hin, isl, ita, jpn, mar, nld, pan, rus, swe, ukr, urd, zho.

Want to add or validate a language? It only requires validating 100 templates. Read more about how to contribute here.

Versions

The dataset is generated from the templates in the GitHub repository and published automatically with every release of the multilingual-gsm-symbolic package. Each release is tagged here with the same version (e.g. v0.5.0). The data changes between versions when templates are fixed or languages are validated, so pin the version to make your results reproducible (see Loading the dataset).

Example analysis

As an example of what the dataset can be used for, we evaluated openai/gpt-5.4-nano (reasoning effort: low) with inspect-ai on English and Danish. These results were produced in April 2026 on an earlier version of the dataset, so they are not directly comparable to evaluations on later versions.

Language	Original accuracy	Synthetic accuracy
English	90.0%	75.2%
Danish	83.2%	70.2%

The gap between original and synthetic accuracy reflects performance degradation on novel number combinations — a proxy for how much a model relies on memorisation vs. genuine reasoning.

Performance degradation (original → synthetic)

Language gap (English vs. Danish, synthetic)

Accuracy by problem difficulty (reasoning steps)

Eval logs are available in the logs/ folder of this repository.

Citation

If you use this dataset, please cite the Multilingual GSM-Symbolic paper:

@misc{enevoldsen2026multilingualgsmsymbolicdeterminescapability,
      title={Multilingual GSM-Symbolic: What determines capability transfer across languages?}, 
      author={Kenneth Enevoldsen and Riley Herchert and Sofie Mosegaard and Dan Saattrup Smart and Simon Enni and Isaac Chung and Sofie Bruun and Ayush Sunil Munot and Max Müller-Eberstein and Adnan El-Assadi and Elisa Bassignana and Gianluca Barmina and Hafsteinn Einarsson and Iben Nyholm Debess and Linda Freienthal and Lukas Galke Poech and Mike Zhang and Nicolas Legrand and Vladimir Salnikov and Yevhen Kostiuk and Zafar Hussain and Sagandeep Kaur and Agnes Toftgård and Marie Mattson and Kristoffer Nielbo},
      year={2026},
      eprint={2610.03367},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2610.03367}, 
}

---

A Danish Foundation Models dataset