davidanugraha/GPT-5.6-Sol-FreeController-Qwen3.5-35B-A3B-SWE-Smith-Step1

GPT-5.6-Sol FreeController / Qwen3.5-35B-A3B / SWE-Smith — step 1

Public portability snapshot for the continual-learning SWE-Smith campaign.

• —Policy optimizer step: 1
• —Campaign state at upload: Wave 7 completed; next coordinator turn blocked by a duplicate local wave-index resume defect
• —Base revision: 59d61f3ce65a6d9863b86d2e96597125219dc754
• —LoRA rank/alpha: 32/64
• —Context length: 65,536
• —Training backend: veRL replay, world size 8, sequence parallel size 8
• —veRL revision: 8f8b122195ecf43475679fb6ee0204ee33f387c5
• —Adapter directory digest: 6b806b4b9c4aacafa6eac1510ce2d79280a783326822c83adb1b6a16e9d74fbd
• —FSDP checkpoint digest: 40f71ef4191a3967d325f10303364d3a47b0038fafc9004bc99a28c59debb057

The adapter files at repository root are sufficient for inference when loaded over the exact base revision. Exact optimizer continuation requires training_checkpoint/global_step_1, the matching code revision, and the separate public campaign-state repository.