Dataset Card for the EIFFEL benchmark

EIFFEL (Evaluation of Idiomatic French Fixed Expressions for Large Language Models) is a French evaluation dataset designed to assess large language models’ (LLM) knowledge of idiomatic expressions in context. EIFFEL comprises 602 samples in multiple choice format designed to test an LLM's capacity to complete an idiomatic expression in context. The samples have been manually constructed and annotated by native speakers and linguists.

• Curated by: This dataset has been built by LINAGORA, OpenLLM-France & l’IRIT (Institut de recherche en Informatique de Toulouse).
• Funded by: OpenLLM France (Bpifrance), ANITI (ANR-19-PI3A-0004) and LLM4All (ANR-23-IAS1-0008).
• Computing resources: provided by GENCI at IDRIS through the grant 2025-AS011016445.
• Language (NLP): French
• License: cc-by-sa-4.0

Dataset Details

Dataset Description and Curation Rationale

The dataset, as explained in the paper EIFFEL: a novel benchmark to measure bias of English heavy training on French idiomatic expressions, contributes to a larger body of work aimed at testing the impact of anglocentric training on LLM performance in non-English languages, especially performance on culturally-specific tasks involving content that is difficult to translate.

In order to target our research question about anglocentricity, we organized the dataset based on whether a given French idiomatic expression has a (direct) translation in English. While some idiomatic expressions can be translated word-by-word between French and English, such as “Not my cup of tea/Pas ma tasse de thé” others are less direct or even completely different. The expression “call a spade a spade” for example, has an obvious counterpart in French, literally “call a cat a cat” but it is not a direct translation. The majority of expressions are even less easily translatable. “Avoir du chien” literally, “to have some dog” means to be charming.

Idiomatic expressions in EIFFEL are categorized accordingly:

• word-for-word (88 expressions): French idioms that have an exact translation in English.
• similar (100 expressions): French idioms for which there is a corresponding English idiomatic expression that is almost a translation of the French expression except for one significant difference, usually a noun or verb phrase.
• different (414 expressions): French idioms for which we have been unable to find English translations or that involve multiple significant differences (multiple noun phrases or noun and verb phrases, for example) from the nearest English expression.

Dataset Structure

EIFFEL is designed as a cloze task such that for each French idiomatic expression in EIFFEL, we mask an important part of the expression and provide four possible ways of filling in the mask. For example, in "Boire comme un templier" (literally, "Drink like a Knight of Templar"), we mask "templier" and provide the options: "templier", "chevalier" (knight), "dauphin" (dolphin) and "poisson" (fish). We also create a context for each expression that encourages an idiomatic interpretation of the target expression as some masked expressions, like "That's not my cup of <...>", could have valid non-idiomatic completions that are not our focus here.

The dataset thus contains the following fields for each sample:

• French with context: a contextual sentence that includes the idiomatic expression in use, with one term in the expression masked.
• French: the target French idiomatic expression.
• English: where possible, the English translation of the French idiomatic expression (work in progress)
• masked sentences: the French idiomatic expression with a term masked.
• answer A - answer D: four options for filling in the masked portion of the masked sentence.
• answer: the correct answer (A - D).
• type: word by word, similar or different.
• register: some expressions are annotated as vulgar.
• Data Origin: whether we intially found the expression in French and subsequently searched for a translation in English or vice versa.

Example :

{
      'French with context': 'Il a une sacrée descente, il boit comme un < ...>.', 
      'French': 'Boire comme un templier', 
      'English': 'To drink like a fish', 
      'masked sentences': 'Boire un comme un < ...>', 
      'answer A': 'templier', 
      'answer B': 'chevalier', 
      'answer C': 'dauphin', 
      'answer D': 'poisson', 
      'answer': 'A', 
      'type': 'similar', 
      'register': None, 
      'Data Origin': 'from French to English'
     
 }

Dataset Creation

For a detailed description of the annotation pipeline, see Appendix E of EIFFEL: a novel benchmark to measure bias of English heavy training on French idiomatic expressions.

Example Use

from datasets import load_dataset

ds_diff = load_dataset("OpenLLM-France/EIFFEL", "different", split="train")
ds_sim = load_dataset("OpenLLM-France/EIFFEL", "similar", split="train")
ds_wbw = load_dataset("OpenLLM-France/EIFFEL", "word by word", split="train")

Bias, Risks, and Limitations

In compiling this dataset, we focused on idiomatic expressions primarily used in mainland France, although some may also be used in other French-speaking regions. It would be worthwhile to expand EIFFEL to include idiomatic expressions specific to other French-speaking cultural areas, such as Belgium, Switzerland or Quebec.

While the dataset is manually annotated and the annotations have been reviewed multiple times, EIFFEL may still contain errors or at least debatable annotations. In general, problems can be traced to the difficulty of translating idiomatic expressions. Translation, as explained in Appendix E of EIFFEL: a novel benchmark to measure bias of English heavy training on French idiomatic expressions, is crucial for determining the category of each expression and in many cases, choosing the mask and distractors.

LLM Evaluation

We provide results for select LLMs on the EIFFEL benchmark for comparison. These LLMs can be divided into three categories:

• Models with at least 30% French training data: Croissant 1.3b; Lucie 7B; Luciole 1b, 8b and 23b; Gaperon 1b, 8b, and 24b
• Models that we assume are anglocentric (over 90% English and code data): Llama 3.2 1b, Llama 3.1 8b, Llama 3.1 70b; Gemma 2 9b, gemma-3-1b-pt,
• Models that fall between these two categories with less that 50% English data but smaller proportions (e.g. 5-10%) of French data: Apertus 70b; EuroLLM-1.7B, EuroLLM 9b.

In the tables below "W-W" stands for "word-by-word", "Sim" for "similar, and "Diff" for "different".

70B models	Avg	W-W	Sim	Diff
Apertus 70B	.95	.95	.94	.96
Llama 70B	.92	.93	.90	.93
Luciole 23b	.97	.98	.97	.97
Gaperon 24b	.95	0.95	0.95	0.94

8B models	Avg	W-W	Sim	Diff
Gaperon 8b	.95	.97	.94	.95
Lucie 7b	.94	.97	.91	.95
Luciole 8b	.96	.98	.95	.96
Eurollm 9b	.93	.97	.90	.93
Llama 3.1 8b	.89	.94	.86	.86
Gemma 9b	.89	.97	.81	.89

1B models	Avg	W-W	Sim	Diff
Croissant 1.3b	.95	.97	.92	.95
Luciole 1b	.93	.97	.89	.93
Gaperon 1b	.92	.95	.89	.92
Eurollm 1.7b	.85	.93	.78	.82
Llama 3 1b	.66	.80	.59	.59
Gemma 3 1b	.78	.86	.72	.76

As described in the full paper, we also trained seven 1 billion parameter bilingual models on varying proportions of French and English data to further investigate the impact of data proportions on LLM performance.

The results for each model are given below, where the first column indicates the percentage of French data used in training.

Proportion fr	Avg	W-W	Sim	Diff
100%	.89	.91	.87	.91
66%	.91	.90	.90	.92
50%	.89	.91	.86	.89
33%	.85	.88	.82	.86
5%	.71	.78	.67	.66
1%	.50	.57	.47	.46
0%	.35	.36	.30	.38

The models can be found at bilingual-french-english-ablation-models.

Citation

When using the EIFFEL Dataset, please cite the following paper (accepted at ACL 2026; more details to follow):

✍ Charlotte Noel, Nicholas Asher, Olivier Gouvert, Farah Benamara, Julie Hunter (2026). EIFFEL: a novel benchmark to measure bias of English heavy training on French idiomatic expressions. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).

@inproceedings{openllm2025Eiffel,
      title={EIFFEL: a novel benchmark to measure bias of English heavy training on French idiomatic expressions}, 
      author={Charlotte Noel and Nicholas Asher and Olivier Gouvert and Farah Benamara and Julie Hunter},
      year={2026},
      booktitle={Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
      url={https://hal.science/hal-05619209}, 
}