BART is a transformer encoder-decoder (seq2seq) model with a bidirectional (BERT-like) encoder and an autoregressive (GPT-like) decoder. BART is pre-trained by (1) corrupting text with an arbitrary noising function and (2) learning a model to reconstruct the original text.

This model is particularly effective when fine-tuned for text generation tasks (e.g., summarization, translation) but also works well for comprehension tasks (e.g., text classification, question answering).

Training details

Dataset: Spanish-legal-corpora - 90% for training / 10% for validation.
Training script: see here

Evaluation metrics 🧾

Metric	# Value
Accuracy	0.86
Loss	0.68

Benchmarks 🔨

WIP 🚧

How to use with `transformers`

from transformers import BartForConditionalGeneration, BartTokenizer

model_id = "mrm8488/bart-legal-base-es"

model = BartForConditionalGeneration.from_pretrained(model_id, forced_bos_token_id=0)
tokenizer = BartTokenizer.from_pretrained(model_id)

def fill_mask_span(text):
  batch = tokenizer(text, return_tensors="pt")
  generated_ids = model.generate(batch["input_ids"])
  print(tokenizer.batch_decode(generated_ids, skip_special_tokens=True))

text = "Los españoles son <mask> ante la ley."
fill_mask_span(text)
# Output: ['Los españoles son iguales ante la ley.1.ª y 2.ª ante la']

text = "Los proyectos de reforma Constitucional deberán <mask> por una mayoría de tres quintos de cada una de las Cámaras."
fill_mask_span(text)
# Output: ['Los proyectos de reforma Constitucional deberán ser aprobados por una mayoría de tres quintos de cada']

Acknowledgments

Citation

If you want to cite this model, you can use this:

bibtex

@misc {manuel_romero_2023,
	author       = { {Manuel Romero} },
	title        = { bart-legal-base-es (Revision c33ed22) },
	year         = 2023,
	url          = { https://huggingface.co/mrm8488/bart-legal-base-es },
	doi          = { 10.57967/hf/0472 },
	publisher    = { Hugging Face }
}

Created by Manuel Romero/@mrm8488

Made with ♥ in Spain

mrm8488/bart-legal-base-es

BART Legal Spanish ⚖️