LegoX: open-source agentic data pipelines, training and evals
I'm Cavan Tao, co-founder of LegoX. LegoX is a series of open-source projects covering agentic data pipelines, training frameworks, evaluation benchmarks and foundation models. Our aim is to connect these pieces into reproducible, extensible systems. Recent releases include Lego-RL, a framework for harness-native reinforcement learning of coding agents, and LegoFlow, with LegoFlow-SWE offering 5,000 open-source tasks and 2,780 successful trajectories filtered from 12M PRs.
What's working: our SWE-Lego models and datasets passed 200K downloads on Hugging Face. What we're figuring out: how coding-agent RL generalizes across harnesses, namely OpenHands SDK, Claude Code and OpenCode, and whether mixed-harness training beats single-harness training or model merging.
I'd love to meet anyone who trains or evaluates coding agents and would be willing to try our releases and give me candid feedback.
Comments1
Welcome to Lounge, Cavan. Filtering 12M PRs down to 2,780 successful trajectories is a steep cut, so I'm curious what drove most of the attrition. On the harness question, do you expect mixed-harness training to hold up better when the tool schemas differ a lot, like Claude Code versus OpenHands?
1 comment
Welcome to Lounge, Cavan. Filtering 12M PRs down to 2,780 successful trajectories is a steep cut, so I'm curious what drove most of the attrition. On the harness question, do you expect mixed-harness training to hold up better when the tool schemas differ a lot, like Claude Code versus OpenHands?