V0.5.0 of PSSA: a non-transformer language model

Repository navigation #

First tagged release, matching the crate version (0.5.0).

Highlights across the project so far:

• Training core: parameter-matched transformer baseline, shared-weight Ouro looping, stacked-depth training, batched independent-sequence lanes, resumable learning-rate horizons, gradient clipping and episodic memory value caps, resumable memory growth.
• GPU backends: native CUDA/cuBLAS alongside WebGPU, resident weight matrices, packed kernel dispatch for the SSM and memory stages, parity and timing checks, and refusal to run training on software adapters.
• Qwen distillation: Qwen3-to-PSSA converter and checkpoint emit, a pure-Rust teacher forward, CE + temperature-KL loss verified by finite-difference tests, per-phase timing breakdowns, and unembedding calibration.
• TUI: full restyle (braille charts, retro panels) through the phase 4 through 12c feature set, then a simplification round: one key grammar, five category menus replacing the 27-tab strip, plain-language help and status lines (519 tests green at that merge).
• Docs and site: README rebuilt around measured results and split into focused guides (contributing, donate, model math, Qwen distillation math), new logo, FUNDING.yml, Hugging Face checkpoint pointers.
• Reliability: hardened CLI parsing and remote input bounds, divergent-loss halt before checkpoint save, honest TUI telemetry panels, Windows platform fixes, and guarded dream replay.