Rhumeet/sentimentiq-weights

SentimentIQ — Trained Model Weights

Trained weights for all 10 models from SentimentIQ — a benchmark comparing classical ML through fine-tuned BERT for 5-class sentiment prediction on 7 million Yelp reviews.

Files

File	Model	Test Macro F1
BERT_finetuned.pt	Fine-tuned BERT (best overall)	0.703
LSTM_multilayer.pt	Multi-layer bidirectional LSTM	0.678
Seq2Seq_LuongAttention.pt	Seq2Seq + Luong attention	0.662
Seq2Seq_NoAttention.pt	Seq2Seq, no attention	0.661
GRU_multilayer.pt	Multi-layer bidirectional GRU	0.661
Seq2Seq_BahdanauAttention.pt	Seq2Seq + Bahdanau attention	0.660
LogisticRegression.pkl	TF-IDF + Logistic Regression	0.641
RNN_multilayer.pt	Multi-layer bidirectional RNN	0.637
LinearSVM.pkl	TF-IDF + Linear SVM	0.631
NaiveBayes.pkl	TF-IDF + Multinomial Naive Bayes	0.559

Usage

.pt files are PyTorch state_dict checkpoints — load them into the corresponding model class defined in the training notebook. .pkl files are pickled scikit-learn estimators, loadable directly via pickle.load().

Full Project

See the SentimentIQ GitHub repository for the complete pipeline: EDA, preprocessing, training code, TOPSIS-based model ranking, and error analysis.