English-to-French translation with T5
Fine-tuning T5-small on OPUS Books with mixed-precision training.
Google’s T5-small fine-tuned on the OPUS Books English–French parallel corpus, built with HuggingFace Transformers and the Trainer API.
Setup
- Prompt-based tokenization in T5’s text-to-text format
- FP16 mixed-precision training
- Custom BLEU metric computation via the
evaluatelibrary
Results
- BLEU improved by 61%, from 0.038 to 0.061
- Validation loss reduced from 2.14 to 1.47 over three epochs
Those BLEU numbers are low in absolute terms, which is the honest result for T5-small fine-tuned briefly on a literary corpus — OPUS Books is stylistically far from the model’s pretraining distribution, and three epochs is not enough to close that gap. The relative improvement is the meaningful signal.