Fine-tuning large language models affects machine translation instruction following
Fine-Tuning LLMs for Translation: General Forgetting Mitigation Does Not Preserve MT-Specific Instruction Following
Computation and LanguageMachine Learning
Summary
Fine-tuning large language models (LLMs) with translation data can improve how well they translate but may cause them to forget other abilities. The authors studied if methods that prevent forgetting in general also keep the models good at following translation-specific instructions, like controlling formality or length. They found that common methods help preserve general skills but do not keep translation instructions working well. Only mixing in special control examples during training helped keep some of these translation instructions, but it didn’t work well on new, unseen instructions.
What this means in practice
- •For machine translation engineers: Improve machine translation fine-tuning by selecting techniques that preserve specialized instruction following capabilities for formality and grammatical control.
- •For natural language processing platform teams: Develop fine-tuning workflows that incorporate control-task data mixing to maintain translation modification instructions within customized language models.
Authors
Niklas Scholz, David Thulke, Abdallah Nasir, Will Allred, Evgeny Matusov, Hermann Ney
Abstract
Fine-tuning large language models on parallel data improves translation quality but can cause catastrophic forgetting. Mitigation methods are generally evaluated by retention on general benchmarks. We ask whether these findings transfer to machine translation (MT) fine-tuning and to MT-specific instruction following (MT-IF): instructions that modify a translation, such as formality, grammatical gender, and length control. We compare methods anchored to auxiliary data, to model outputs, and to the base model parameters, first in a screening study with Llama 3.2 1B Instruct, then on Llama 3.1 8B Instruct fine-tuned on bidirectional Arabic-English or Spanish-English data. Elastic Weight Consolidation preserves general capabilities best in both stages; on the 8B Spanish model the average score on general benchmarks drops 1.7 points versus 11.0 for standard fine-tuning, yet its scores for formality and grammatical gender control remain close to standard fine-tuning. Only data mixing with control-task examples preserves these controls, but its gains do not transfer to unseen prompts for the same task.