Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CroissantLLM: A Truly Bilingual French-English Language Model

Manuel Faysse, Patrick Fernandes, Nuno M. Guerreiro, António Loison, Duarte M. Alves, Caio Corro, Nicolas Boizard, João Alves, Ricardo Rei, Pedro H. Martins, Antoni Bigata Casademunt, François Yvon, André F. T. Martins, Gautier Viaud, Céline Hudelot, Pierre Colombo

TL;DR

CroissantLLM tackles English-dominant bias by delivering a truly bilingual 1.3B model trained on a balanced English-French corpus with a specialized bilingual tokenizer. It introduces FrenchBench for targeted evaluation and openly releases data, code, checkpoints, and chat/translation models to empower reproducible research and industrial deployment. Through scaling-law guided data-ratio analysis and a strong emphasis on transparency, the work demonstrates competitive French and cross-language performance for a compact model, while prioritizing on-device inference. Overall, CroissantLLM advances multilingual LLM research and provides a practical, open platform for exploring bilingual pretraining and deployment at modest compute scales.

Abstract

We introduce CroissantLLM, a 1.3B language model pretrained on a set of 3T English and French tokens, to bring to the research and industrial community a high-performance, fully open-sourced bilingual model that runs swiftly on consumer-grade local hardware. To that end, we pioneer the approach of training an intrinsically bilingual model with a 1:1 English-to-French pretraining data ratio, a custom tokenizer, and bilingual finetuning datasets. We release the training dataset, notably containing a French split with manually curated, high-quality, and varied data sources. To assess performance outside of English, we craft a novel benchmark, FrenchBench, consisting of an array of classification and generation tasks, covering various orthogonal aspects of model performance in the French Language. Additionally, rooted in transparency and to foster further Large Language Model research, we release codebases, and dozens of checkpoints across various model sizes, training data distributions, and training steps, as well as fine-tuned Chat models, and strong translation models. We evaluate our model through the FMTI framework, and validate 81 % of the transparency criteria, far beyond the scores of even most open initiatives. This work enriches the NLP landscape, breaking away from previous English-centric work in order to strengthen our understanding of multilinguality in language models.

CroissantLLM: A Truly Bilingual French-English Language Model

TL;DR

CroissantLLM tackles English-dominant bias by delivering a truly bilingual 1.3B model trained on a balanced English-French corpus with a specialized bilingual tokenizer. It introduces FrenchBench for targeted evaluation and openly releases data, code, checkpoints, and chat/translation models to empower reproducible research and industrial deployment. Through scaling-law guided data-ratio analysis and a strong emphasis on transparency, the work demonstrates competitive French and cross-language performance for a compact model, while prioritizing on-device inference. Overall, CroissantLLM advances multilingual LLM research and provides a practical, open platform for exploring bilingual pretraining and deployment at modest compute scales.

Abstract

We introduce CroissantLLM, a 1.3B language model pretrained on a set of 3T English and French tokens, to bring to the research and industrial community a high-performance, fully open-sourced bilingual model that runs swiftly on consumer-grade local hardware. To that end, we pioneer the approach of training an intrinsically bilingual model with a 1:1 English-to-French pretraining data ratio, a custom tokenizer, and bilingual finetuning datasets. We release the training dataset, notably containing a French split with manually curated, high-quality, and varied data sources. To assess performance outside of English, we craft a novel benchmark, FrenchBench, consisting of an array of classification and generation tasks, covering various orthogonal aspects of model performance in the French Language. Additionally, rooted in transparency and to foster further Large Language Model research, we release codebases, and dozens of checkpoints across various model sizes, training data distributions, and training steps, as well as fine-tuned Chat models, and strong translation models. We evaluate our model through the FMTI framework, and validate 81 % of the transparency criteria, far beyond the scores of even most open initiatives. This work enriches the NLP landscape, breaking away from previous English-centric work in order to strengthen our understanding of multilinguality in language models.
Paper Structure (49 sections, 12 figures, 15 tables)

This paper contains 49 sections, 12 figures, 15 tables.

Table of Contents

  1. Introduction
  2. Contributions
  3. Data
  4. French Data
  5. English Data
  6. Code Data
  7. Parallel Data
  8. Training
  9. Model Architecture
  10. Tokenizer
  11. Selecting an optimal language ratio
  12. Final data distribution
  13. Training framework
  14. Training losses
  15. Environmental impact
  16. ...and 34 more sections

Figures (12)

  • Figure 1: Conversation example with CroissantLLMChat
  • Figure 2: Fertility on unseen test sets using various tokenizers. Lower is better.
  • Figure 3: Evolution of test cross-entropy loss with model size in English (left) and French (right), for the wiki domain, as well as the fitted joint scaling laws,
  • Figure 4: Effective capacity ratio (as predicted by our fitted joint scaling law) for English and French as we change the weight of each language.
  • Figure 5: (Left) Training loss with respect to the number of seen tokens. (Right) Validation perplexity (Averaged Log Likelihood) on Wikitext (English), computed with a rolling stride
  • ...and 7 more figures