Skip to content

Improve wandb-primary training diagnostics adoption - #3210

Open
morganmcg1 wants to merge 3 commits into
mainfrom
codex/wandb-diagnostics-adoption-audit
Open

Improve wandb-primary training diagnostics adoption#3210
morganmcg1 wants to merge 3 commits into
mainfrom
codex/wandb-diagnostics-adoption-audit

Conversation

@morganmcg1

Copy link
Copy Markdown
Member

Summary

  • audit recent conversation logs for wandb-primary training-diagnostics adoption
  • document the findings in analysis/wandb_primary_training_diagnostics_adoption_audit_2026-04-23.md
  • tighten the skill/reference guidance around step_key, completion criteria, and phrase-to-helper mapping
  • add a one-shot curve_diagnostics_cli.py wrapper to lower the activation energy for actually running the helper workflow

Findings

  • scanned the available 2026-04-21 and 2026-04-22 logs (137 archives / 1,388 embedded .jsonl transcripts)
  • found 41 assistant-side wandb-primary skill launches
  • found 0 assistant-side invocations of the concrete diagnostics helper surface (plot_single_run_overview, compare_runs_curves, plot_run_comparison, curve_features, lr_schedule_features, grad_norm_features, grad_histogram_features, list_candidate_step_keys, guess_step_key_from_workspace) after excluding injected skill text
  • found strong heuristic-level adoption instead: step-axis corrections, LR-peak spike analysis, plateau vs convergence distinctions, overfit reasoning, and stability-aware run ranking
  • found a new runtime blocker: the detected uv environment currently lacks pandas, even though the diagnostics helper stack imports it at module import time

Why This May Help

  • lower activation energy: the new CLI turns the default workflow into one command
  • stronger completion gate: the skill now pushes toward step_key + helper-derived table + PNG + verdict instead of manual narration alone
  • better mapping from natural research phrases like gradient storm and trough envelope plateaued to the helper surface that should back those claims
  • stronger step-axis and restart guidance, based on concrete failure modes seen in the logs

Validation

  • uv run python -m py_compile .agents/skills/wandb-primary/scripts/curve_diagnostics_cli.py
  • uv run python .agents/skills/wandb-primary/scripts/curve_diagnostics_cli.py --help
  • uv run python .agents/skills/wandb-primary/scripts/curve_diagnostics_cli.py --entity foo --project bar --run baz (expected helpful missing-dependency message for pandas in the current uv environment)

mmcguire and others added 3 commits April 23, 2026 13:16
- CLAUDE-STUDENT.md: add metric-curve analysis requirement to run reports
- icml2026/airfrans/program.md: clarify paper-facing scorer normalization
- icml2026/drivaerml/program.md: correct AB-UPT / Transolver lineage
- icml2026/train.py: parity/robustness tweaks
- uv.lock: dep updates
- .gitignore: exclude large local scratch dirs (conversation_logs, .claude/worktrees, analysis/operational_metrics, analysis/scratch, tm)

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants