Godlydonuts/Sai

Sai Initiative Sai is Project Shohin's return to its original objective: build the strongest practical model near four billion parameters. This repository is the live scratchpad and implementation surface for that effort. Nothing is called an improvement until it beats the unchanged parent and an equal-compute control on real, source-disjoint benchmarks. Data precedes architecture. Sai first earns a trustworthy learning sequence: verified source bytes, quality and duplication… See the full description on the dataset page: https://huggingface.co/datasets/Godlydonuts/Sai.