ShadowPEFT – Centralized and Detachable Parameter-Efficient Fine-Tuning
Detachable PEFT modules that version independently, unlike LoRA's weight-coupled adapters.

Training code withheld during patent review limits reproducibility and community adoption.
ML researchers, NLP engineers studying model interpretability
LoRA · QLoRA · AdapterHub
Detachable PEFT modules that version independently, unlike LoRA's weight-coupled adapters.
178K parameter model beats a Pokemon roguelike after the author rage-quit hundreds of times.
Native ternary training beats post-training quantization for memory efficiency.
Karpathy's microgpt in C99, proves tiny coordinated models beat single large models on logic.
Bulk export Perplexity conversations to Obsidian Markdown with Spaces organization and deduplication.
High-recall search API for event monitoring that beats Exa on benchmarks.