STSB DistilRoBERTa Base

cross-encoder/stsb-distilroberta-base is a text-ranking cross-encoder model that computes a semantic similarity score between 0 and 1 for pairs of sentences. It is hosted on gigarouter as a managed, OpenAI-compatible API for reranking tasks.

Model architecture and training

Built on the distilroberta-base backbone (82 million parameters), the model was fine-tuned on the SentenceTransformers STS benchmark dataset (sentence-transformers/stsb). It is English-only and licensed under Apache 2.0.

Performance and use

The model accepts two sentences and returns a single similarity score, making it suitable for semantic reranking pipelines where candidate passages must be scored against a query. Its output is calibrated for the STS benchmark, a widely used evaluation for semantic textual similarity.

With over 4.5 million total downloads and support for PyTorch, JAX, ONNX, and OpenVINO, the model has been validated across environments. ONNX and quantized variants are available, enabling efficient deployment.

As a cross-encoder, it directly compares both sentences through a transformer, producing more accurate relevance scores than bi-encoder alternatives at the cost of higher latency per pair. Gigarouter hosts the model in an optimized inference stack, so developers benefit from low-latency API calls without managing infrastructure.