The Fine-Tuning Paradox: Boosting Translation Quality Without Sacrificing LLM Abilities

2024/05/30 by David Stap, Eva Hasler, Stap, David +7 · 1 voice · 7 citations
Arts and Humanities · Chemistry · Computer Science · Psychology · #Artificial intelligence #Boosting (machine learning) #Chemistry #Computation and Language (cs.CL) #Computer science #Epistemology #FOS: Computer and information sciences #Fine-tuning #Natural Language Processing Techniques #Philosophy #Physics #Psychology #Quality (philosophy) #Translation (biology) #Translation Studies and Practices #cs.CL

paper · pdf · doi:10.48550/arxiv.2405.20089

published in arXiv (Cornell University) (Cornell University)

openalex publication_date 2024/05/30 · arxiv published 2024/05/30 · openalex created_date 2024/06/01 · arxiv updated 2024/08/06 · openalex updated_date 2026/07/28

Abstract

Fine-tuning large language models (LLMs) for machine translation has shown improvements in overall translation quality. However, it is unclear what is the impact of fine-tuning on desirable LLM behaviors that are not present in neural machine translation models, such as steerability, inherent document-level translation abilities, and the ability to produce less literal translations. We perform an extensive translation evaluation on the LLaMA and Falcon family of models with model size ranging from 7 billion up to 65 billion parameters. Our results show that while fine-tuning improves the general translation quality of LLMs, several abilities degrade. In particular, we observe a decline in the ability to perform formality steering, to produce technical translations through few-shot examples, and to perform document-level translation. On the other hand, we observe that the model produces less literal translations after fine-tuning on parallel data. We show that by including monolingual data as part of the fine-tuning data we can maintain the abilities while simultaneously enhancing overall translation quality. Our findings emphasize the need for fine-tuning strategies that preserve the benefits of LLMs for machine translation.

Cited by

• Hearing to Translate: The Effectiveness of Speech Modality Integration into LLMs
• Isolating Culture Neurons in Multilingual Large Language Models
• RL from Teacher-Model Refinement: Gradual Imitation Learning for Machine Translation
• It's Not a Walk in the Park! Challenges of Idiom Translation in Speech-to-text Systems
• Compensating for Data with Reasoning: Low-Resource Machine Translation with LLMs
• The Effect of Language Diversity When Fine-Tuning Large Language Models for Translation
• "Be My Cheese?": Cultural Nuance Benchmarking for Machine Translation in Multilingual LLMs
• Two Intermediate Translations Are Better Than One: Fine-tuning LLMs for Document-level Translation Refinement

Discussions

• LLMs just are surpisingly bad at maintaining the information and tone of what they translate, you can effectively only have one or the other at the moment [bsky, 1 points, 1 comments]

Related

• How Multilingual Are Large Language Models Fine-Tuned for Translation?
• Fine-Tuning Large Language Models to Translate: Will a Touch of Noisy Data in Misaligned Languages Suffice?
• A Novel Paradigm Boosting Translation Capabilities of Large Language Models
• Lost in Literalism: How Supervised Training Shapes Translationese in LLMs
• Translation in the Wild