Sparsity Improves Ternary Language Models: Evidence from BitNet b1.58

This deposit contains the artifacts and documentation for the RPT project, focused on efficient language models with ternary weights (BitNet b1.58), sparsity, and QAT/STE. The main validated result improves PPL from 25.13 to 16.39 (-34.8%) with 42.6% sparsity and 100% ternary weights, with functional CPU deployment via GGUF i2_s. It also includes notebooks and scripts used to reproduce experiments, as well as logs and validation documentation. Repository and references:- Code and paper: https://bit.ly/GitHubRPT- Model: https://bit.ly/rpt-bitnet-2b-pruned- GGUF: https://bit.ly/rpt-bitnet-2b-pruned-GGUF Note: arXiv submission is in progress (pending endorsement/moderation).