Show HN: Cbjev – typed decisions about text from one encoder pass

Category: ai-ml

Tags: nlp, encoder, text-classification, pytorch, inference-optimization, typed-decisions

Score: 7.5/10 (Innovation: 7, Technical: 8, Documentation: 8, Utility: 7)

Cbjev is a Python library for efficient typed decisions (choice, score, noul) over text or JSON using a single encoder pass. It packs multiple questions into one forward pass, achieving significant speedups over Laya while maintaining or improving accuracy and calibration. It offers a drop-in replacement for Laya with a custom ModernBERT implementation, CUDA graph optimizations, and multilingual support.

Target audience: machine learning engineers, data scientists, backend developers integrating NLP

Repository: https://github.com/tomek7667/cbjev · Python · GPL-3.0 · 5 stars

View on Hacker News