HelsinkiNLP-FineTuned-Legal-es-zh

This model is a fine-tuned version of Helsinki-NLP/opus-tatoeba-es-zh on a dataset of legal domain constructed by the author himself.

Intended uses & limitations

This model is the result of the master graduation thesis for the Tradumatics: Translation Technologies program at the Autonomous University of Barcelona.

Please refer to the GitHub repo created for this thesis for the full-text and relative open-sourced materials: https://github.com/guocheng98/MUTTT2020_TFM_ZGC

The thesis intends to explain various theories and certain algorithm details about neural machine translation, thus this fine-tuned model only serves as a hands-on practice example for that objective, without any intention of productive usage.

Training and evaluation data

The dataset is constructed from the Chinese translation of Spanish Civil Code, Spanish Constitution, and many other laws & regulations found in the database China Law Info (北大法宝 Beida Fabao), along with their source text found on Boletín Oficial del Estado and EUR-Lex.

There are 9972 sentence pairs constructed. 1000 are used for evaluation and the rest for training.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 2000
• num_epochs: 10
• mixed_precision_training: Native AMP
• weight_decay: 0.01
• early_stopping_patience: 8

Training results

Best validation loss achieved at step 5600.

Framework versions

• Transformers 4.7.0
• Pytorch 1.8.1+cu101
• Datasets 1.8.0
• Tokenizers 0.10.3