ALECTRA-small-OWT

This is an extension of ELECTRA small model, trained on the OpenWebText corpus. The training task (discriminative LM / replaced-token-detection) can be generalized to any transformer type. Here, we train an ALBERT model under the same scheme.

Pretraining task

(figure from Clark et al. 2020)

ELECTRA uses discriminative LM / replaced-token-detection for pretraining. This involves a generator (a Masked LM model) creating examples for a discriminator to classify as original or replaced for each token.

The generator generalizes to any *ForMaskedLM model and the discriminator could be any *ForTokenClassification model. Therefore, we can extend the task to ALBERT models, not just BERT as in the original paper.

Usage

from transformers import AlbertForSequenceClassification, BertTokenizer

# Both models use the bert-base-uncased tokenizer and vocab.
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
alectra = AlbertForSequenceClassification.from_pretrained('shoarora/alectra-small-owt')

NOTE: this ALBERT model uses a BERT WordPiece tokenizer.

Code

The pytorch module that implements this task is available here.

Further implementation information here, and here is the script that created this model.

This specific model was trained with the following params:

batch_size: 512
training_steps: 5e5
warmup_steps: 4e4
learning_rate: 2e-3

Downstream tasks

GLUE Dev results

Model	# Params	CoLA	SST	MRPC	STS	QQP	MNLI	QNLI	RTE
ELECTRA-Small++	14M	57.0	91.	88.0	87.5	89.0	81.3	88.4	66.7
ELECTRA-Small-OWT	14M	56.8	88.3	87.4	86.8	88.3	78.9	87.9	68.5
ELECTRA-Small-OWT (ours)	17M	56.3	88.4	75.0	86.1	89.1	77.9	83.0	67.1
ALECTRA-Small-OWT (ours)	4M	50.6	89.1	86.3	87.2	89.1	78.2	85.9	69.6

GLUE Test results

Model	# Params	CoLA	SST	MRPC	STS	QQP	MNLI	QNLI	RTE
BERT-Base	110M	52.1	93.5	84.8	85.9	89.2	84.6	90.5	66.4
GPT	117M	45.4	91.3	75.7	80.0	88.5	82.1	88.1	56.0
ELECTRA-Small++	14M	57.0	91.2	88.0	87.5	89.0	81.3	88.4	66.7
ELECTRA-Small-OWT (ours)	17M	57.4	89.3	76.2	81.9	87.5	78.1	82.4	68.1
ALECTRA-Small-OWT (ours)	4M	43.9	87.9	82.1	82.0	87.6	77.9	85.8	67.5