Cleo: trying to fit full analyst behavior in a 2B model [P]
A new open-source project named Cleo has been developed, aiming to integrate full analyst behavior into a small, 2-billion parameter model. This project fine-tunes a Qwen3.5-2B-Base model and emphasizes a unified harness for training, evaluation, and inference. Key features include training on data preparation and answering contracts, searching queries with live execution evidence, and co-designing the model's safety, SQL handling, and clarification behaviors as a single system. All components, including the harness, model, and datasets, are publicly available. AI
IMPACT This project demonstrates the feasibility of integrating complex analyst behaviors into smaller models, potentially enabling more efficient and accessible AI solutions for data analysis tasks.