OpenMOSS/Say-I-Dont-Know

Provides model-specific "I don't know" (Idk) datasets and multiple fine-tuning approaches—Idk-SFT, Idk-BoN with reward modeling, Idk-DPO, and Idk-PPO—to align LLMs to refuse unanswerable questions while maintaining knowledge on familiar topics. Built on llama-recipes, DeepSpeed-Chat, and DPO frameworks with FSDP distributed training support across models like Llama-2, Baichuan2, and Mistral. Includes inference tools for both prompt-based and learned refusal strategies, transforming Unknown Unknowns and Unknown Knowns into calibrated Known states via knowledge quadrant classification.

No commits in the last 6 months.