dusersad12/SST2-Sentiment-Analysis-Results

SST2 Sentiment Analysis Results

---

Overview

This dataset card presents the results of our sentiment analysis experiments on the SST-2 (Stanford Sentiment Treebank) benchmark. We fine-tuned BERT-base with various hyperparameter configurations to find the optimal setup for binary sentiment classification.

Experimental Setup

All experiments used BERT-base-uncased as the backbone model. We explored different learning rates, batch sizes, dropout rates, and weight decay values. Each configuration was trained on the SST-2 training set and evaluated on the validation set.

Best Run Results

The best performing configuration achieved the following metrics on the SST-2 validation set:

Metric	Value
eval_accuracy	0.924
eval_f1	0.921
eval_precision	0.933
eval_recall	0.910
eval_loss	0.234
train_loss	0.145

Best Run Configuration

Hyperparameter	Value
learning_rate	5e-05
batch_size	32
epochs	5
dropout	0.10
weight_decay	0.01

Comparison Across Runs

Run	Learning Rate	Batch Size	Epochs	Dropout	F1 Score
run_001	2e-05	32	3	0.10	0.858
run_002	5e-05	16	3	0.10	0.888
run_003	1e-04	32	5	0.20	0.869
run_004	3e-05	64	3	0.15	0.904
run_005	5e-05	32	5	0.10	0.921
run_006	2e-05	16	5	0.20	0.895
run_007	1e-04	64	3	0.30	0.851
run_008	3e-05	16	5	0.15	0.909

Citation

If you use these results, please cite:

@misc{sst2-sentiment-analysis,
  title={SST-2 Sentiment Analysis Experiment Results},
  author={Research Team},
  year={2025}
}

License

This work is licensed under the Apache-2.0 License.