laion/jtd-d4-34-30B

Jupiter TaskTrove DAPO campaign, arm D4 — DAPO + partial-credit arm — D1 + Harbor threshold reward shaper (1.0/0.3), filter on shaped rewards. Dataset: laion/codeforces-v2 (10,000 competitive-programming tasks; late steps trained on the TaskTrove codeforces-v3 snapshot after laion/codeforces-v2 was deleted). Base model Qwen/Qwen3-Coder-30B-A3B-Instruct; FSDP2 fully-async trainer, 16 policy GPUs (4 nodes x 4 GH200) + 4 vLLM engines (TP2); Harbor/Daytona sandboxed terminus-2 rollouts.

• Checkpoint: step 34 of 80, selected by trailing-5 EMA of reward/avg_raw_reward over the arm's full step series (campaign ended early on platform degradation + data-quality grounds; final banked step 35).
• Launch config: rl_config.yaml in this repo; metrics + per-link logs in training_logs/.
• pass@8 reads 1.0 on the dynamic-sampling arms (d1/d2/d4) are post-filter by construction; use reward/avg_raw_reward series in training_logs/ for cross-arm comparison.

Training Traces

Training-time Daytona/Harbor rollouts: penfever/jtd-d4 — the last episode of each trial, conserved at a 25% subsample: every 4th trial dir of the arm's full trace store (33231 enumerated at upload; 167 shards, 21824 rows). Owner-instructed conservation quota for this campaign's cleanup, not the uploader default.