| name | ieee-rl-dl-ml-reproducible-training |
| description | Use when RL/DRL/DL/ML paper reproduction or energy-system control work needs unattended long-goal execution, long training, full-resume checkpoints, mandatory TensorBoard long-run visualization when applicable, traceable train/test evaluation, same-tier raw environment reward delta comparisons when applicable, online-checkable CSV/JSONL/stdout records, IEEE TSG/TPS figures, Chinese-first colleague-readable reports, post-training PPT/report packs, figure/table quality audit, SHA256 manifests, artifact indexes, risk analysis, or cleanup/exclusion of failed or half-finished artifacts. |
IEEE RL/DL/ML Reproducible Training
Overview
Use this skill to keep RL/DRL/DL/ML reproduction work traceable end to end: live records, separate train/test evaluation when applicable, IEEE-style figures, Chinese-first colleague-readable reports, risk analysis, reproducibility manifests, and clean final experiment folders.
Scope Guard
If project instructions declare the work as non-RL, affine-only, DL-only, or paper-style numerical reproduction, reuse only the recordkeeping, IEEE figure, manifest, cleanup, and final-audit parts of this skill. Do not impose RL-only artifacts such as TensorBoard events, rewards, episodes, policy mode, train/eval rollouts, alpha/lambda traces, or feasible-rate metrics unless the project explicitly defines them.
Reference Routing
| User task | Load |
|---|
| Set up or audit unattended long-goal execution, long training traces, full-resume checkpoints, scientific gates, five-cycle debugging, subagent monitors/reviewers/safeguards, stdout logs, CSV/JSONL, TensorBoard, checkpoints, smoke tests, or evaluation/test passes | references/realtime-training-monitoring.md |
Write or audit README, requirements, output.md, run records, environment records, or Python headers | references/reproducibility-recordkeeping.md plus needed template in assets/ |
| Generate PPT-ready post-training Markdown, figures, tables, reproducibility manifest, artifact index, PPT index, or missing-figure/table reports | references/post-training-reporting.md; also load references/ieee-plot-style.md for figure style |
| Generate or review IEEE Transactions on Smart Grid / Power Systems figures, captions, export formats, SVG fonts, or figure manifests | references/ieee-plot-style.md; use scripts/check_ieee_plot_manifest.py when a manifest exists |
| Final audit before declaring a run/project/report complete | references/final-quality-gates.md; optionally run scripts/audit_reproducible_training_project.py |
| Reviewer-driven closure for substantial code, experiment, plotting, reporting, or metric-definition changes | references/final-quality-gates.md |
| Clean or close a project stage, exclude failed/half-finished/junk artifacts, or preserve successful runs | references/project-hygiene-cleanup.md |
| Use MATLAB for live CSV viewing | references/realtime-training-monitoring.md and assets/monitor_training_template.m |
| Verify an existing run folder | Use scripts/validate_run_trace.py |
Do not load every reference by default. Load only the routed files plus any directly required assets/scripts. This section is the hard router.
Hard Gates
- Do not start long training without a smoke test, resource/GPU/Python/solver check when relevant, dry-run manifest, unique output directory, live CSV/JSONL/stdout/checkpoint plan, and TensorBoard event logging/dashboard plan.
- Use subagents during long training only when explicitly requested or already authorized by the project plan; keep them as bounded reviewers, monitors, and safeguards while the main agent owns scientific decisions, debug patches, run stops, cleanup, and final claims.
- Do not call a run resumable unless checkpoints include full training state: model, target model, optimizer, scheduler/noise state, replay/buffer state when applicable, RNG states, episode/global-step cursor, config hash, git commit, and resume command. Actor-only checkpoints are warm starts, not full resume.
- Do not start an unattended repository long goal without recording
git status --short --branch, reading AGENTS.md, README.md, relevant plan/index/protocol docs, configs, tests, and target source files, and identifying the first incomplete required gate.
- Do not claim training/reporting completion until train and evaluation records are separated and the final quality gates are checked.
- For substantial code, experiment, plotting, reporting, or metric-definition changes, run the reviewer-driven closure loop in
references/final-quality-gates.md; unresolved P0/P1 findings block completion.
- Do not advance a phase after a failed scientific gate. Run the Brooks-style five-cycle debug rule from
references/realtime-training-monitoring.md, then mark BLOCKED with evidence if still unresolved.
- Post-training reward comparisons must use same-tier raw environment reward delta, such as environment
step() reward or eval_episodes.csv:reward on one scale. Do not compare shaped training-objective rewards across methods as performance claims.
- Root project
README.md must be a navigation page with relative links to reports, figures, tables, code, configs, runs, validation, manifests, and missing-output notes; do not leave it as only an introduction or changelog.
- Generated Markdown must be Chinese-first, technically explicit, cross-linked, and readable by an engineering colleague who did not watch the run.
- Post-training packs must include README or index entry, main/summary/debug reports, risk analysis, figure/table index, reproducibility manifest with SHA256 and bytes, and explicit missing-output notes.
- Figures must follow IEEE style, export PNG/PDF/SVG when data exists, validate SVG font handling and vector path geometry, avoid overlap/clutter, and pass a figure quality audit or record why audit is unavailable.
- Before editing generated Python experiment files, load
assets/python_file_header_templates.md. All generated .py files keep the Chinese-first overview header fields without a literal # 中文为主总览: line. Buddha ASCII is an additional block only for main.py, run_*.py, train_*.py, long-running launchers, and equivalent project entrypoints. Important or innovative code blocks need concise Chinese comments.
- Preserve successful run records. Delete failed/misleading artifacts only when user/project instructions authorize deletion; otherwise mark them excluded with evidence.
- Missing data is a result. Do not fabricate plots, tables, SOC traces, metrics, baselines, or paper-level claims.
Minimal Workflow
- Define the minimum reproducible path: hypothesis, algorithm, environment, seeds, steps, metrics, and outputs.
- Before long training, create real-time records and run a smoke test.
- During long training, write live CSV/JSONL, stdout summaries, TensorBoard events, and checkpoints; use MATLAB CSV monitoring only as the backup live viewer.
- After training, run a separate evaluation/test pass with explicit checkpoint, policy mode, test seeds, and metrics.
- After valid training/evaluation packages, generate PPT-ready reports, risk notes, figure/table packages, figure quality audit, manifests, and artifact indexes, or explicit missing-artifact notes.
- Run the final quality gates, then clean or exclude misleading artifacts according to the project cleanup policy; keep trustworthy successful runs.
Operating Rules
- Do not wait until a run ends to write artifacts.
- Treat CSV/JSON artifacts as the authoritative reproducibility record. TensorBoard is the required long-training live dashboard; MATLAB CSV viewing is the backup supervision path, not a reason to run end-to-end blind.
- Keep train and test records separate.
- Treat IEEE style as scientific communication, not decoration.
- Keep generated project documentation Chinese-first with bilingual technical terms at first mention.
- Do not invent generic training termination rules. Use project-defined criteria when the project provides them.
- Do not treat training as complete until post-training reporting artifacts are generated, cross-linked, quality-audited, or missing artifacts are documented.
- Do not delete run artifacts unless the user or project instructions explicitly authorize cleanup. If deletion is not authorized, mark failed or misleading outputs as excluded.
- Preserve successful multi-seed, multi-algorithm, and repeated validation runs with time/seed/script naming.
- Do not create extra documentation inside this skill folder. Project
README.md, requirements.txt, and output.md are outputs guided by the skill, not skill-internal README files.
Verification Commands
Run only the checks that match existing artifacts:
python scripts/validate_run_trace.py path/to/run-folder
python scripts/check_ieee_plot_manifest.py path/to/figure_manifest.json
python scripts/audit_reproducible_training_project.py --project-root path/to/repo --results-dir path/to/results
After editing this skill itself, validate the skill folder structure and markdown formatting.