BiLSTM (ESIM-style) for RTE
Trained on RTE with frozen GloVe-300d embeddings.
Validation accuracy: 0.5783. Trainable params: 967,170.
Files: best_model.pt (state_dict), vocab.json, config.json.
Load with the ESIMLite class (hid=128), passing a dummy embedding matrix of shape (vocab_size, 300).
