ConvBERT base pre-trained on large_spanish_corpus

The ConvBERT architecture is presented in the "ConvBERT: Improving BERT with Span-based Dynamic Convolution" paper.

Metrics on evaluation set

disc_accuracy = 0.9488542
disc_auc = 0.8833056
disc_loss = 0.15933733
disc_precision = 0.79224133
disc_recall = 0.27443287
global_step = 1000000
loss = 9.658503
masked_lm_accuracy = 0.6177698
masked_lm_loss = 1.7050561
sampled_masked_lm_accuracy = 0.5379228

Usage

from transformers import AutoModel, AutoTokenizer
model_name = "mrm8488/convbert-base-spanish"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModel.from_pretrained(model_name)

Created by Manuel Romero/@mrm8488 with the support of Narrativa

Made with <span style="color: #e25555;">♥</span> in Spain