OpenThoughtsAgent: Open Data Recipes for Training Terminal and Software Engineering Agents

AI Technology: Virtual agents; Natural Language Processing

Agents such as Claude Code are becoming increasingly powerful for solving  software engineering problems, such as fixing an issue in a codebase, or installing a complicated software environment. 

The critical component of an agentic system is the large language model (LLM) that orchestrates commands, for example which commands in a terminal to execute next. The critical ingredient for building a LLM that performs well in an agentic system is the training data, in particular the fine-tuning data. There is currently insufficient knowledge on how to curate or generate high-quality fine-tuning data to enable agentic systems. 

This project aims to close this gap by developing a testbed and a framework for creating very strong open-source agents (as measured on TerminalBench and SWEBench) for autonomously completing multi-step tasks in a terminal environment. We will study and develop the best design principles for data pipelines for improving model abilities to interact with its environment through supervised fine-tuning and reinforcement learning. 

The study will be making systematic interventions to the training data and open-source the full end-to-end pipeline for data curation, model training, environments, tools and evaluation to foster the progress in this crucial field of modern machine learning. The project contributes to advanced AI produced in the EU.