Fine-Tuning LLMs for Translation: General Forgetting Mitigation Does Not Preserve MT-Specific Instruction Following
Elastic Weight Consolidation limits general forgetting in translation fine-tuning but not translation-specific instruction following.
The authors test whether general catastrophic-forgetting mitigations preserve machine-translation instruction following, including formality, grammatical gender, and length control. Screening used Llama 3.2 1B Instruct, followed by Llama 3.1 8B Instruct fine-tuned on bidirectional Arabic-English or Spanish-English data. Elastic Weight Consolidation preserved general benchmarks best: on the 8B Spanish model, the average general score fell 1.7 points versus 11.0 for standard fine-tuning, while formality and gender scores stayed close to standard fine-tuning. Only mixing in control-task examples preserved those controls, and the gains did not transfer to unseen prompts for the same task.
- Study covers formality, grammatical gender, and length-control instructions.
- Tests use Llama 3.2 1B and Llama 3.1 8B Instruct.
- EWC cut an 11.0-point general-score drop to 1.7 on 8B Spanish.
- Only control-example mixing preserved controls, and not on unseen prompts.
Full article159 words · extracted from arxiv.org · click to collapse
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.
Text extracted automatically; images, tables and formatting may be missing. Original: https://arxiv.org/abs/2609.28395