TinyBERT is 7.5 times smaller and 9.4 times faster on inference compared to its teacher BERT model (while DistilBERT is 40% smaller and 1.6 times faster than BERT). This model was trained using the entire dataset (3.6M of samples) in constrast to the distilbert model which was trained on only 10% of the dataset.

Intended uses & limitations

While this model may not be as accurate as the distilbert model, its performance should be enough for most use cases.

python

from transformers import pipeline

# Create the pipeline
sentiment_classifier = pipeline('text-classification', model='AdamCodd/tinybert-sentiment-amazon')

# Now you can use the pipeline to classify emotions
result = sentiment_classifier("This product doesn't fit me at all.")
print(result)
#[{'label': 'negative', 'score': 0.9969743490219116}]

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 3e-05
train_batch_size: 32
eval_batch_size: 32
seed: 1270
optimizer: AdamW with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
lr_scheduler_warmup_steps: 150
num_epochs: 1
weight_decay: 0.01

Framework versions

Transformers 4.35.0
Pytorch lightning 2.1.0
Tokenizers 0.14.1

If you want to support me, you can here.