laion/jtd-d0-69-30B

GRPO control checkpoint (step 69 of 80; the run ended early at step 72 after its restart chain exhausted) from the Jupiter TaskTrove DAPO campaign, arm D0 — plain GRPO, symmetric clipping (eps 0.2/0.2), token-level loss mean, no dynamic sampling, no reward shaping.

• Base model: Qwen/Qwen3-Coder-30B-A3B-Instruct
• Training data: laion/codeforces-v2 (10,000 competitive-programming tasks, sandboxed agent rollouts via Harbor/Daytona, terminus-2 harness, FSDP2 fully-async GRPO on 16× GH200)
• Checkpoint selection: trailing-5 EMA of reward/avg_raw_reward (0.2755 at step 69 vs 0.2354 at step 72)
• Launch config: rl_config.yaml in this repo
• Metrics + logs: training_logs/ in this repo (per-link metrics CSVs, trial results, report)

Training Traces

Training-time Daytona/Harbor rollouts: penfever/jtd-d0 (the last episode of each trial — the rollouts the policy trained on after rollback/truncation).