sentiment-bert-base

Fine-tuned BERT-base for binary sentiment classification on the Sentiment140 dataset (1.6M tweets).

Base model

google-bert/bert-base-uncased — the original BERT-base-uncased from Devlin et al. (2019), 110M parameters.

Training

• Dataset: Sentiment140 (1.6M tweets, 80/20 split, seed 42)
• Hyperparameters: learning rate 2e-5, batch size 16, 3 epochs
• Hardware: NVIDIA A10G, AWS SageMaker (g5.2xlarge)
• Training time: 7.3 hours
• Trainer: Hugging Face Transformers + Trainer API; load_best_model_at_end=True

Test set performance

Metric	Value
Accuracy	87.46%
Precision	0.880
Recall	0.869
F1	0.874

Intended use

Demonstration model for an academic purposes

Limitations

• English only, binary sentiment, 2009-era Twitter language.
• Sentiment140 labels generated automatically using emoticons (distant supervision), introducing systematic noise.
• Does not handle sarcasm reliably (the dataset does not separate it as a phenomenon).