LegoFlow-SWE pushes Qwen3.5 to 70.2%

The real story is data selection: carefully verified tasks can outperform simply scaling trajectory volume. The result is promising, but benchmark and anti-leakage controls matter as much as the headline score.

• –Only 5,000 tasks survived filtering from the original 12M-PR pool, underscoring how scarce genuinely useful coding data remains.
• –LegoFlow-SWE includes 2,780 successful trajectories, giving developers reusable training material rather than just evaluation examples.
• –Under the same setup, LegoFlow-SWE improves over SWE-rebench V2 by 5.8 points on SWE-bench Verified.
• –LegoFlow’s Curator, Tracer, Trainer, and Evaluator blocks make the full data-to-model pipeline more reproducible.
• –The 70.2% score reflects a controlled benchmark setup, not guaranteed production-level software reliability.