tt-x0_verifier-shaped-20-30B

tt-x0_verifier-shaped (TaskTrove X0 shaped verifier, global_step_20)

X0 verifier-ablation shaped (pass_ratio) arm of the TaskTrove hyperparameter sweep. Completed its 20-step horizon. Base: Qwen/Qwen3-Coder-30B-A3B-Instruct; terminus-2 agentic RL on DCAgent/exp_rpt_multifile (pytest verifier, pass_ratio reward shaping). Trailing-5 EMA of the shaped reward at step 20 = 0.5094 (inflated by partial credit vs the binary arm's 0.1703; not cross-comparable).

X0 verdict (owner 2026-08-04): SHAPED WINS on the post-#300 unshaped re-score (pass@k 0.347 vs 0.310, ~3.7-pt gap). pass_ratio shaping is the campaign verifier for X1-X5. See training_logs/report.md.

Training Traces

Not available. The trial-level trace trees were reclaimed during a 2026-08 GPFS inode-recovery pass before a dataset was uploaded to the Hub; no tt-x0_verifier-shaped-traces dataset was published and the source trials no longer exist on GPFS. Per-trial unshaped verifier outcomes survive on Jupiter in /e/scratch/jureap59/feuer1/x0_unshaped_outcomes.tsv. The training_logs/ analysis here is derived from the per-step WANDB mirror, not the (deleted) trials.