nllb-200-distilled-600M-OpenSLR-GL-EN

This model is a fine-tuned version of facebook/nllb-200-distilled-600M on juanjucm/OpenSLR-SpeechT-GL-EN datasetfor Galician-to-Englis Machine Translation task. It takes Galician texts as input and generates the correspondant English translation.

This Machine Translation model, was developed to be the second stage of a Speech Translation cascade system for transcribing and translating Galician audios into English texts. This STT model can be used as a first step to transcribe Galician audio into text. After that, this MT model can be applied over the generated Galician transcriptions to get English text translations.

The motivation behind this work is to increase the visibility of the Galician language, making it more accessible for non-Galician speakers to understand and engage with Galician audio content.

This model was developed during a 3-week Speech Translation workshop organised by Yasmin Moslem.

Performance and training details

Baseline model achieved a BLEU score of 51.32 on the evaluation dataset.

After fine-tuning, it achieves the following results on the evaluation set:

• Loss: 0.0122
• BLEU: 73.6259

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 8
• optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• num_epochs: 8

Training results

We used BLEU Score as our reference translation metric for selecting the best checkpoint after training.

Training Loss	Epoch	Step	Validation Loss	Bleu
14.2627	1.0	600	3.7799	61.8432
6.0125	2.0	1200	0.5403	66.7094
1.1534	3.0	1800	0.0243	69.1604
0.0748	4.0	2400	0.0147	70.7523
0.0125	5.0	3000	0.0131	73.1040
0.0095	6.0	3600	0.0126	73.2385
0.0081	7.0	4200	0.0122	73.8670
0.0072	8.0	4800	0.0122	73.6259

Framework versions

• Transformers 4.47.1
• Pytorch 2.4.1+cu121
• Datasets 3.2.0
• Tokenizers 0.21.0