Experiment: Increase embedding weight decay from 0.0005 to 0.001 to test whether stronger regularization on token embeddings improves validation bpb, following the success of the lm_head weight decay increase.

Campaign

• Campaign: regularization-tuning

Hypothesis

Increase embedding weight decay from 0.0005 to 0.001 to test whether stronger regularization on token embeddings improves validation bpb, following the success of the lm_head weight decay increase.

Parent Context

• Parent master hash: 935fdbf9f4ae8a5ef5bcb76552acea2bc5801965
• Master val_bpb at dispatch: 0.962777
• Worker id: worker-0
• Worktree: /Users/ben/code/open-autolab/.runtime/worktrees/embedding-wd-001

Single Variable

• train.py line 257: weight_decay = 0.001 (was 0.0005) for embedding optimizer group

Expected Upside

Stronger regularization on token embeddings could improve generalization and validation bpb, following the success of the lm_head weight decay increase.

Duplicate Check

This is not a duplicate - no previous experiments have tested embedding weight decay changes on the current master.

Runtime

• Log path: /Users/ben/code/open-autolab/research/live/embedding-wd-001.log
• Launcher: uv run scripts/opencode_worker.py run embedding-wd-001

Allowed Edit Scope

• train.py only

Run Plan

• Refresh master with uv run scripts/refresh_master.py --fetch-dag
• Run uv run scripts/hf_job.py preflight
• Run uv run scripts/hf_job.py launch --mode experiment
• Stream logs to the reserved path
• Parse uv run scripts/parse_metric.py /Users/ben/code/open-autolab/research/live/embedding-wd-001.log

Result

• Status: CANCELLED (STUCK)
• HF Job: 69cbdff7942f980bf425a3d0
• Stuck at: step 02348 (97.6%) during final evaluation phase
• Duration: 10+ minutes hanging before cancellation
• Submitted: no
• Interpretation: Job cancelled due to infrastructure issue during evaluation. Hypothesis remains untested.
• Failure mode: Infrastructure - job hung during evaluation phase

Memory-Keeper Handoff

• One short note for research/notes.md: CANCELLED due to job hanging at step 02348 during evaluation. Hypothesis remains untested.
• Any do-not-repeat update: Add to Known Regressions under Regularization Tuning section with note about infrastructure issue.