EverydayGPT-205M
Official model associated with the paper:
EverydayGPT: Confidence-Gated Routing for Efficient and Safe Hybrid GPT-RAG Conversational QA
Author: Jaspreet Singh Nahal
Overview
EverydayGPT is a lightweight GPT-based language model designed for use within a Confidence-Gated Routing (CGR) architecture. The model serves as the generation component of a hybrid GPT-RAG system that dynamically routes queries between retrieval, generation, and refusal pathways.
The research focuses on efficient CPU deployment, safe retrieval augmentation, and confidence-based routing under resource constraints.
Associated Research
The full methodology, evaluation protocol, and results are described in:
EverydayGPT: Confidence-Gated Routing for Efficient and Safe Hybrid GPT-RAG Conversational QA
arXiv:2606.11212
Model Details
- Architecture: GPT-style Transformer
- Vocabulary: GPT-2 BPE (50,257 tokens)
- Training Corpus: FineWeb-Edu
- Intended Use: Research and experimentation
- Language: English
Repository Contents
- model checkpoint
- configuration files
- tokenizer assets (if available)
Citation
If you use this model, please cite:
Jaspreet Singh Nahal. EverydayGPT: Confidence-Gated Routing for Efficient and Safe Hybrid GPT-RAG Conversational QA. arXiv:2606.11212.
