AbstractQbit/electra_large_imdb_htsplice

google/electra-large-discriminator finetuned on imdb dataset for 2 epoches.

Large examples tokenized with head and tail parts of a review, as described in How to Fine-Tune BERT for Text Classification?

def preprocess_function(example):
    tokens = tokenizer(example["text"], truncation=False)
    if len(tokens['input_ids']) > 512:
        tokens['input_ids'] = tokens['input_ids'][:129] + \
            [102] + tokens['input_ids'][-382:]
        tokens['token_type_ids'] = [0]*512
        tokens['attention_mask'] = [1]*512
    return tokens