GRPO checkpoint from the TaskTrove X4 (rollout temperature) sweep. Base model Qwen/Qwen3-Coder-30B-A3B-Instruct, trained on DCAgent/exp_rpt_multifile with SkyRL + Terminus-2; campaign verifier is pass_ratio shaping.

global_step_45 was selected by the trailing-5 reward EMA (alpha = 1/3) over the full restart chain -- the highest-EMA aligned checkpoint (EMA 0.2328; step reward 0.2305; pass@8 0.375; entropy 0.078), per parse_skyrl_metrics.py --run_dir --save_every 5.

The temp-1.0 arm finished both dataset epochs (step 66, matching sibling arm temp 0.7) with a healthy profile: policy entropy declined ~0.25x (0.21 -> 0.05) -- no entropy runaway. The post-horizon extension (to max_steps 80) was stopped by the owner at step 68. Exports 55-65 were lost to infra chain-stalls before saving; step 45 is the best saved export (the FSDP checkpoint banked at 66 but was not HF-exported).

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

open-athena/tt-x4_temperature-t1p0 -- 1/4 systematic subsample (every 4th trial, uniform coverage of the full 66-step horizon; 15,258 rows from 68,768 trials). The full set was GPFS-read-bound on the Jupiter login node (~6h); subsample is a documented, owner-approved deviation.