cea-list-ia/MathAleaQA

MathAléaQA

MathAléaQA is a French dataset of 342,691 math question–solution pairs covering the French school curriculum from CM1 (primary school) to terminale (final year of high school). It is extracted from MathALÉA, the open-source exercise generator developed by the Coopmaths association. Each example has a question, the worked solution written by the exercise author, and a short final answer extracted by an LLM.

The dataset is meant for supervised fine-tuning (SFT) and for open-answer evaluation of LLMs on French school mathematics. It was introduced in Duignan et al. (2026) alongside MathAléaMCQ, Exo7QA and Exo7MCQ.

from datasets import load_dataset

ds = load_dataset("cea-list-ia/MathAleaQA")

Dataset structure

Field	Description
type_id	MathALÉA exercise template identifier (e.g. 1AL11-4, can2F01). All examples that share a type_id are random instances of the same template.
level	French school level: CM1_CM2, sixième, cinquième, quatrième, troisième, seconde, première, terminale, or other.
question	Exercise statement in French, with LaTeX math ($...$).
solution	Worked solution from MathALÉA, in French with LaTeX.
extracted_answer	Short final answer extracted from solution by Qwen3-4B-Instruct-2507 with few-shot prompting. It keeps French number formatting and has not been manually checked.

Example

{
  "type_id": "1AL11-4",
  "level": "première",
  "question": "$(t_n)$ est une suite géométrique de raison $q=-0,8$ et de premier terme $t_1=-5$.\nCalculer $t_{10}$.\nDonner la valeur arrondie au dixième.",
  "solution": "La suite $(t_n)$ est géométrique de raison $q=-0,8$ et de premier terme $t_1=-5$.\nOn en déduit que pour tout $n\\in\\mathbb{N}^*$, \n$\\begin{aligned} t_n&=t_1\\times q^{n-1}\\\\ t_n&=-5\\times (-0{,}8)^{n-1}\\\\ t_{10}&=-5\\times (-0{,}8)^{9} \\end{aligned}$ \nAinsi, $t_{10}\\simeq{0,7}$.",
  "extracted_answer": "0,7"
}

Splits

Split	Examples	Templates (type_id)
train	324,090	1,100
validation	3,273	760
test	15,328	57

The split is made by template, not by example. No type_id appears in both train and test, so the test split measures generalisation to exercise types never seen in training. Validation examples are drawn from the training templates, which makes the validation split in-distribution and suited to early stopping. About 5% of templates go to test and 1% of the remaining examples to validation (seed 42).

Because whole templates are assigned to test, the level mix of the test split differs from train. For example, seconde makes up 41% of test but 22% of train, and test has no other examples:

Statistics

Examples	342,691
Templates (type_id)	1,157
Levels	9
Median question length	90 characters / 13 words
Median solution length	169 characters / 23 words

Examples per level

Level	Examples	%
CM1_CM2	23,278	6.8
sixième	61,239	17.9
cinquième	42,708	12.5
quatrième	21,550	6.3
troisième	43,583	12.7
seconde	79,520	23.2
première	41,353	12.1
terminale	25,096	7.3
other	4,364	1.3

Examples per template

Each template is capped at 1,000 examples so that a few prolific generators cannot dominate. 173 templates reach the cap and together make up 50.5% of the dataset. The other 984 templates have fewer examples.

Question and solution length

	Mean	Median	p90	p99	Max
Question (characters)	106.5	90	206	346	717
Solution (characters)	224.2	169	461	1,066	2,976

Construction

1. Extraction. Non-MCQ exercise blocks (\begin{EXO}) are extracted from MathALÉA's LaTeX exports, giving 351,869 raw question–solution pairs.
2. Cleaning and filtering. LaTeX is cleaned while math environments are kept. The following are removed: exercises that rely on a figure (TikZ, Scratch, coordinate grids), multi-part questions, empty solutions, solutions identical to their question, and duplicate questions. Each type_id is then capped at 1,000 examples, leaving 342,718 pairs (2.6% removed).
3. Answer extraction. Qwen3-4B-Instruct-2507, served with vLLM, extracts a short final answer from each solution using few-shot prompting.
4. Split. The train/validation/test split is made by type_id, as described above. At this step, 27 rows whose extracted_answer is a long LLM error message are dropped, giving the final 342,691 examples.

Intended uses

• Supervised fine-tuning. Use question as the user turn and solution as the assistant turn.
• Open-answer evaluation. Compare a model's final answer with extracted_answer on the test split, whose templates never appear in train. For a curated benchmark with verifiable answers, see MathAleaQA-1K. It holds 1,000 items from this test split whose answers are canonicalised and checked with math-verify.

Limitations

• Template-generated. Examples of the same type_id share wording and structure and mostly differ in their numbers. The effective diversity is closer to 1,157 templates than to 342K independent problems.
• No figures. Exercises that need a figure were removed, so geometry is under-represented.
• extracted_answer is model output. It is unverified, can include units or French formatting (0,7, \dfrac{3}{4}), and is sometimes wrong. Normalise it before exact-match scoring, or use MathAleaQA-1K.
• Short solutions. Solutions are MathALÉA's concise corrections, not step-by-step reasoning traces.
• French curriculum. Levels follow the French school system.

License

The exercises come from MathALÉA by Coopmaths. Coopmaths distributes its educational content under CC BY-SA, and this dataset is released under the same license (CC BY-SA 4.0). The MathALÉA source code is licensed separately under GNU AGPLv3.

Citation

@inproceedings{duignan:hal-05708459,
  title     = {Jeux de donn{\'e}es en fran{\c c}ais pour l'affinage et l'{\'e}valuation de mod{\`e}les de langue g{\'e}n{\'e}ratifs dans le domaine des math{\'e}matiques},
  author    = {Duignan, Liam and Graiess, Asma and van Ypersele, Matteo and Deshayes-Chossart, J{\'e}r{\^o}me and Ferret, Olivier},
  booktitle = {Actes de l'atelier Intelligence Artificielle et {\'E}DUcation (IA EDU)},
  editor    = {Badache, Ismail},
  publisher = {ATALA},
  address   = {Nantes, France},
  pages     = {74--97},
  year      = {2026},
  month     = jun,
  url       = {https://hal.science/hal-05708459},
  hal_id    = {hal-05708459}
}

Acknowledgements

We thank the teachers of the Coopmaths association, who develop MathALÉA and publish its exercises under an open license.