Yes, they do not curate new GitHub issues to be fair. But I think they curated the trajectories themselves using the Qwen3 Coder 480B variant. Basically this is just SWE Gym on steroids (not any method novelty). Also, I am somewhat concerned that they did not filter the repositories from SWE Bench Verified, but (repo, commit) tuples and issue descriptions:
https://x.com/togethercompute/status/2026738066321453408. So it may have happened that they trained on the same repo but from a different commit id if they were not careful enough