tt-x1_lr-lr4e6 (TaskTrove X1 lr=4e-6, global_step_30)
X1 learning-rate ablation arm (lr = 4e-6) of the TaskTrove hyperparameter sweep.
- Base model:
Qwen/Qwen3-Coder-30B-A3B-Instruct - Harness: terminus-2 agentic RL
- Source:
DCAgent/exp_rpt_multifile(TaskTrove, pytest verifier,pass_ratioreward shaping) - Geometry: 6 nodes x 4 GH200, Jupiter (JSC)
Status
TERMINATED by owner at step 36/80 after reward collapsed to exactly 0.000 at steps 33-36 (a hard
collapse, not the gradual decay expected from an over-large learning rate). This is not a
completed / valid X1 result: it did not reach its 80-step horizon. Selected checkpoint =
global_step_30, the trailing-5 EMA peak (0.1700) among saved exports and the last healthy
checkpoint before the step-33 collapse. See training_logs/report.md for the full curve.
Training Traces
Companion trial-level trace dataset: https://huggingface.co/datasets/open-athena/tt-x1_lr-lr4e6
training_logs/
Per-step metric surface, the parse_skyrl_metrics.py analysis report, and the reward plot
(metrics.csv, report.md, reward_plot.png).