laion/jtd-d5-69-30B

GSPO checkpoint (step 69; the run closed at step 73 of 80 when its restart chain was sealed at close-out — the final links ran through the 2026-08-28 GPFS/Daytona platform degradation) from the Jupiter TaskTrove DAPO campaign, arm D5 — the D0 plain-GRPO control with the objective swapped to GSPO: policy_loss_type: gspo, eps 3e-4/4e-4 (paper range), loss_reduction: sequence_mean, use_tis: false, no dynamic sampling, no KL.

• 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 trainer on 16x GH200)
• Checkpoint selection: trailing-5 EMA of reward/avg_raw_reward (0.2784 at step 69 vs 0.2481 at step 72; the same step wins on EMA of reward/avg_pass_at_8, 0.4845 vs 0.4723)
• Launch config: rl_config.yaml in this repo
• Metrics + logs: training_logs/ in this repo (per-link metrics CSVs, trial results, report, all 17 job logs + the checkpoint-export job log)

Training Traces

Training-time Daytona/Harbor rollouts: penfever/jtd-d5 (the last episode of each trial — the rollouts the policy trained on after rollback/truncation). This trace dataset conserves a deterministic 25% subsample of the full trial set (17,722 of 70,886 trials; every 4th trial of the sorted list) = 15,880 rows / 93 shards, per a storage-conservation decision at cleanup time.