Muennighoff/sgpt

Implements both bi-encoder and cross-encoder architectures for GPT models, using parameter-efficient fine-tuning (BitFit) on bias tensors and position-weighted mean pooling for symmetric/asymmetric search tasks. Integrates with Hugging Face Transformers and Sentence Transformers, with pre-trained models optimized on MS MARCO and NLI datasets evaluated across BEIR and USEB benchmarks. The newer GritLM successor unifies bi-encoders, cross-encoders, and generative capabilities in a single model.

873 stars. No commits in the last 6 months.