Yanagi-Origami/autocode-rl-gptoss20b-synthetic

autocode-rl-gptoss20b-synthetic

GPT-OSS-20B after full-parameter GRPO against a deterministic execution judge, trained on 147 frontier-written rewrites of real Codeforces problems.

• —Expert weights are stored fused (mlp.experts.gate_up_proj, 411 tensors), the layout the official gpt-oss release uses and the only one SGLang's disk loader understands. The per-expert layout that Megatron exporters produce loads without error and then emits gibberish, so it is deliberately not what is published here.
• —Endpoint checkpoint of a step budget fixed in advance: no selection on any evaluation set.
• —Evaluation protocol, and the judge, are described in the accompanying paper.

Pass@1, mean of 4 samples, C++ solutions graded by the benchmark's official tests.