SimCSE

Why we included this project

SimCSE is the reference implementation behind the EMNLP 2021 paper on simple contrastive learning of sentence embeddings. The unsupervised version learns from plain text alone, using standard dropout as noise, while the supervised variant feeds on NLI data, treating entailment pairs as positives and contradictions as hard negatives. Pretrained checkpoints load directly through the bundled sentence-embedding tool or Hugging Face, so you can get usable vectors without retraining anything. Training and evaluation code is included too, which matters if you want to reproduce the reported numbers or adapt the approach to your own data. For teams weighing sentence-embedding options, having the original implementation and its evaluation scripts together beats reimplementing from the paper.