tt-x2_clip-hi0p01 — step 15 (PPO upper-clip = 0.01)

GRPO checkpoint from the TaskTrove X2 (upper PPO clip) hyperparameter sweep. Base model Qwen/Qwen3-Coder-30B-A3B-Instruct, trained on DCAgent/exp_rpt_multifile with SkyRL + Terminus-2. This is the clip-0.01 arm; the campaign verifier is pass_ratio shaping.

Checkpoint selection

global_step_15 was selected by the trailing-5 reward EMA (alpha = 1/3) over the full restart chain — the highest-EMA aligned checkpoint (EMA 0.1642; step reward 0.191; pass@8 0.359; entropy 0.127). hf_save_interval was 5; the first save (step 5) is excluded.

⚠ This run collapsed — not a valid X2 result

The clip-0.01 arm was killed by the entropy stop rule at step 62 (banked 60, horizon 80). Reward flatlined around 0.13–0.22 (pass@8 ~0.19–0.44) with no clear upward trend for steps 1–54, policy entropy climbed monotonically (0.11 at step 1 -> 0.75 at step 50), then entropy blew up to 3.0 and reward/pass@8 crashed to ~0 at steps 55–62. Step 15 is the EMA-best checkpoint, taken before the entropy-driven decline; later checkpoints (50/55/60) are post-decline or post-collapse. This arm did not reach its horizon and is not a valid X2 sweep result on the same terms as the killed X1 lr-4e-6 arm.

See training_logs/ for metrics.csv, report.md, reward_plot.png, the resolved rl_config.json, and the gzipped .out chain.

Training Traces

Training-time Daytona/Harbor rollouts: open-athena/tt-x2_clip-hi0p01 (the last episode of each trial — the rollouts the policy trained on after rollback/truncation).