Production-grade, test-driven encoder-only Transformer built from scratch with clean architecture and modular design.
-
Updated
May 10, 2026 - Python
Production-grade, test-driven encoder-only Transformer built from scratch with clean architecture and modular design.
Cross-lingual authorship attribution as a fine-tuned small language model (mT5 encoder). 91.2% on full 705-passage eval. Two HF models, journal, corpus rebuild path.
Retired Cinqic research project exploring an encoder-only routing model. Preserved as a historical research archive; no model was released.
Scaling Laya's encoder models beyond 2K tokens with sparse attention. Pluggable attention variants, benchmarked against decoder scorers.
To associate your repository with the encoder-only topic, visit your repo's landing page and select "manage topics."