mbart-neutralization

This model is a fine-tuned version of facebook/mbart-large-50 on the dataset hackathon-pln-es/neutral-es (https://huggingface.co/datasets/somosnlp-hackathon-2022/neutral-es). It achieves the following results on the evaluation set:

• Loss: 0.0138
• Bleu: 98.5147
• Gen Len: 18.4583

Model description

This model was delivered as part of a class project, in which we used different NLP libraries and resources.

Intended uses & limitations

The aim of this model is detecting gendered-specific structures in Spanish and neutralize them to get a more inclusive text.

Given the computing limitations, there is room for improvement and obtenining more neutral texts.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5.6e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• optimizer: Use OptimizerNames.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: 2

Training results

Training Loss	Epoch	Step	Validation Loss	Bleu	Gen Len
No log	1.0	440	0.0165	98.4773	18.6042
0.223	2.0	880	0.0138	98.5147	18.4583

Framework versions

• Transformers 4.51.2
• Pytorch 2.10.0+cu128
• Datasets 4.0.0
• Tokenizers 0.21.4