tt-x1_lr-lr8e6 -- step 60 (X1 learning rate 8e-6)

GRPO checkpoint from the TaskTrove X1 (learning rate) 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.

Checkpoint selection

global_step_60 was selected by the trailing-5 reward EMA (alpha = 1/3) over the full restart chain -- the highest-EMA aligned checkpoint (EMA 0.1894; step reward 0.1973; pass@8 0.3125; entropy 0.2878), per parse_skyrl_metrics.py --run_dir --save_every 5.

Run status -- terminated mid-horizon (elevated entropy)

Terminated by the owner at step 71/80. Policy entropy climbed ~3.2x (0.11 -> 0.36) -- elevated and rising (entropy stop rule fires at 10x). Step 60 is the best saved checkpoint. Not a horizon result.

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

Training Traces

open-athena/tt-x1_lr-lr8e6 -- 1/4 systematic subsample (every 4th trial, uniform coverage of steps 1-71; 22,169 rows). The full ~96k-trial set was GPFS-read-bound (~9h on the login node); subsample is a documented deviation for this terminated (71/80) arm.