VityaVitalich/bert-tiny-sst2

Bert Tiny for SST2

This model is a fine-tuned version of M-FAC/bert-tiny-finetuned-sst2 on the sst2 dataset. It achieves the following results on the evaluation set:

• Loss: 0.4771
• Accuracy: 0.8280

Usage Example

from transformers import BertTokenizer, BertForSequenceClassification, TrainingArguments, Trainer, DataCollatorWithPadding
import datasets

model = BertForSequenceClassification.from_pretrained('VityaVitalich/bert-tiny-sst2')
tokenizer = BertTokenizer.from_pretrained('VityaVitalich/bert-tiny-sst2')

def create_data(tokenizer):
  train_set = datasets.load_dataset('sst2', split='train').remove_columns(['idx'])
  val_set = datasets.load_dataset('sst2', split='validation').remove_columns(['idx'])

  def tokenize_func(examples):
      return tokenizer(examples["sentence"], max_length=128, padding='max_length', truncation=True)

  encoded_dataset_train = train_set.map(tokenize_func, batched=True)
  encoded_dataset_test = val_set.map(tokenize_func, batched=True)
  data_collator = DataCollatorWithPadding(tokenizer)
  return encoded_dataset_train, encoded_dataset_test, data_collator

encoded_dataset_train, encoded_dataset_test, data_collator = create_data(tokenizer)

training_args = TrainingArguments(
    output_dir='./results',
    learning_rate=3e-5,
    per_device_train_batch_size=128,
    per_device_eval_batch_size=128,
    load_best_model_at_end=True,
    num_train_epochs=5,
    weight_decay=0.1,
    fp16=True,
    fp16_full_eval=True,
    evaluation_strategy="epoch",
    seed=42,
    save_strategy = "epoch",
    save_total_limit=5,
    logging_strategy="epoch",
    report_to="all",
)


trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=encoded_dataset_train,
    eval_dataset=encoded_dataset_test,
    data_collator=data_collator,
    compute_metrics=compute_metrics,
)

trainer.evaluate(encoded_dataset_test)

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 3e-05
• train_batch_size: 128
• eval_batch_size: 128
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 5

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy
0.2313	1.0	527	0.4771	0.8280
0.2057	2.0	1054	0.4937	0.8257
0.1949	3.0	1581	0.5121	0.8177
0.1904	4.0	2108	0.5100	0.8200
0.1879	5.0	2635	0.5137	0.8211

Framework versions

• Transformers 4.34.0.dev0
• Pytorch 2.0.1+cu117
• Datasets 2.14.5
• Tokenizers 0.14.0