coetaur0/ESIM

The model uses a sequential architecture with bidirectional LSTMs and attention mechanisms to compare premise-hypothesis pairs, supporting multiple NLI datasets (SNLI, MultiNLI, Breaking NLI) through modular preprocessing and training pipelines. It integrates with GloVe word embeddings and provides configuration-driven workflows for data preparation, model training with checkpoint resumption, and evaluation against standard benchmarks. The implementation achieves competitive accuracy on SNLI (88.0%) and MultiNLI (76.6-77.0%), matching or exceeding the original paper's reported performance.

374 stars. No commits in the last 6 months.