Llama-GENBA-10B: A Trilingual Large Language Model for German, English and Bavarian
This addresses the problem of language bias for the German NLP community and promotes low-resource languages like Bavarian, though it is incremental as it builds on existing models and methods.
The paper tackles English-centric bias in large language models by developing Llama-GENBA-10B, a trilingual model for German, English, and Bavarian, which achieves strong cross-lingual performance, with its fine-tuned variant surpassing models like Apertus-8B-2509 and gemma-2-9b in Bavarian and outperforming EuroLLM in English.
We present Llama-GENBA-10B, a trilingual foundation model addressing English-centric bias in large language models. Built on Llama 3.1-8B and scaled to 10B parameters, Llama-GENBA-10B is continuously pretrained on 164B tokens (82B English, 82B German, and 80M Bavarian), balancing resources while preventing English dominance. Targeted at the German NLP community, the model also promotes Bavarian as a low-resource language. Development tackled four challenges: (1) curating a multilingual corpus despite Bavarian scarcity, (2) creating a unified tokenizer for English, German, and Bavarian, (3) optimizing architecture and language-ratio hyperparameters for cross-lingual transfer, and (4) establishing the first standardized trilingual evaluation suite by translating German benchmarks into Bavarian. Evaluations show that Llama-GENBA-10B achieves strong cross-lingual performance, with the fine-tuned variant surpassing Apertus-8B-2509 and gemma-2-9b in Bavarian and establishing itself as the best model in its class for this language, while also outperforming EuroLLM in English and matching its results in German. Training on the Cerebras CS-2 demonstrated efficient large-scale multilingual pretraining with documented energy use, offering a blueprint for inclusive foundation models that integrate low-resource languages.