Integrate the PDSS algorithm, a novel framework that enhances local small language models (SLMs) using differentially private protected Chain of Thoughts (Cot) generated by remote LLMs:
Implement InferDPT for privacy-preserving Cot generation.
Support an encoder-decoder mechanism for privacy-preserving Cot generation.
Add prefix trainers for step-by-step distillation and text encoder-decoder training.
Integrate the FDKT algorithm, a framework that enables domain-specific knowledge transfer from LLMs to SLMs while preserving SLM data privacy
Deployment Optimization: support installation of FATE-LLM by PyPi
Major Features and Improvements