Note The artefact of the talk: one captured rollout with exact token ids and logprobs, plus the code that produced it

sergiopaniego/Qwen3-8B-opencode-deepcoder-grpo

Text Generation • 8B • Updated Jul 28 • 48 • 1

Note The policy trained this way, through the agent own loop

Note The competitive-coding tasks the agent solves

Note Dashboard of the run shown in the talk, 0.27 to 0.71 over 10 steps