Cleo tests analyst-like work in a small 2B model

Cleo is an attempt to make a small 2B model act more like a data analyst. It is built by fine-tuning Qwen3.5-2B-Base. The main idea is to train, test, and run the model inside the same structured setup it will use in real work, instead of training it in one setting and using it in another.

Cleo can create several possible SQL queries, run them, and use the live results as evidence when choosing an answer. The system is designed as one package: the model contract, SQL safety checks, database dialect handling, time limits, and behavior for asking clarifying questions are built together. The code, model, and datasets are all open source.

ECHO is also suggested as a method to explore for teams trying reinforcement learning with limited resources.