TeapotLLM- an open-source <1B model for hallucination-resistant Q&A on a CPU

a year ago

Teapot is an open-source small language model (~800 million parameters) optimized for resource-constrained devices like smartphones and CPUs.
It is fine-tuned on synthetic data to reduce hallucinations and focuses on context-based answers.
Teapot supports tasks like Question Answering (QnA), Retrieval-Augmented Generation (RAG), and JSON extraction.
The model is trained to provide conversational answers and resist hallucinations by refusing to answer without sufficient context.
Teapot can perform RAG across multiple documents and extract structured information in formats like JSON.
It includes a library (teapotai) for easy integration into production environments.
Teapot is fine-tuned from flan-t5-large and trained on a ~10mb synthetic dataset.
The model is licensed under MIT and is community-driven, with support available via Discord.

Hasty Briefsbeta