grtninja/qwen-training-workbench-ops
skill-candidates/qwen-training-workbench-ops/SKILL.md
Operate the local teacher-student Qwen training workbench for long-running dataset-factory, staggered continuation batches, and training-agent supervision. Use when advancing or monitoring the private Huihui-to-Qwen3.5-4B pipeline, switching workstation modes, checking batch artifacts, or resuming overnight local training without media-workbench contention.
$ install --global
skillsauthnpx skillsauth add grtninja/skill-arbiter qwen-training-workbench-opsInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
Security Scan Results
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
3
Scanners Passed
9
Scanners in report
Clean
TrivyContainer and dependency vulnerability scanner
95%
Clean
SemgrepStatic code analysis for vulnerabilities
95%
Clean
mcp-scan (Snyk)Model Context Protocol security validation
95%
Skipped
Snyk (dep)Open source security scanning
50%
Skipped
Socket.devSupply chain security analysis
50%
Skipped
VirusTotalMulti-engine malware detection
50%
Skipped
CrowdStrikeAdvanced threat intelligence
50%
Skipped
OSV-ScannerOpen Source Vulnerability database check
50%
Skipped
OWASP Dep-Check
50%
Last scanned: Apr 5, 2026, 2:29 PM246.8s4 files scanned
SKILL.md
- name:
- qwen-training-workbench-ops
- description:
- Operate the local teacher-student Qwen training workbench for long-running dataset-factory, staggered continuation batches, and training-agent supervision. Use when advancing or monitoring the private Huihui-to-Qwen3.5-4B pipeline, switching workstation modes, checking batch artifacts, or resuming overnight local training without media-workbench contention.
Qwen Training Workbench Ops
Use this skill for the private local training workbench that feeds the media stack.
Workflow
- Confirm the workstation is in the correct local mode before touching the lane.
- Identify whether the active step is dataset factory or staggered continuation.
- Check the long-running training supervisor on
127.0.0.1:9041before starting or resuming work. - Validate the required local model lanes:
- teacher lane on a loopback-hosted OpenAI-compatible runtime
- baseline student lane on a loopback-hosted OpenAI-compatible runtime
- embedding lane on a loopback-hosted OpenAI-compatible runtime
- Use artifact progress, not guesswork, to decide whether to continue, wait, or recover.
- If the dataset lane fails, inspect logs and artifacts before any continuation batch is allowed to run.
Phase Routing
Use dataset_factory when the goal is to build or expand the distilled teacher-student dataset:
- weighted Penny-first selection
- Huihui teacher distillation
- MX3 and embedding enrichment
- merged dataset artifact production
Use nsf_staggered_batches when the goal is to keep training the student forward in microbatches:
- batch-level adapter carry-forward
- per-batch
trainer.report.json - per-batch
checkpoint.eval.json - exact source refs in
batch_sources.jsonandbatch_sources.txt
Do not use this skill for:
- Media rendering/export work. Use the media-workbench lane instead.
- Generic LM Studio troubleshooting without training-workbench context.
- Publishing or pushing model artifacts. This lane is private local only.
Canonical Commands
Supervisor status:
curl http://127.0.0.1:9041/v1/training-agent/status
curl http://127.0.0.1:9041/v1/training-agent/jobs/<job_id>
Dataset factory launch:
powershell -ExecutionPolicy Bypass -File `
<training-workbench-root>\tools\start_penny_dataset_factory_nsf_expanded.ps1
Continuation batch launch:
powershell -ExecutionPolicy Bypass -File `
<training-workbench-root>\tools\start_qwen35_4b_radeon_nsf_staggered.ps1
LM Studio teacher check:
lms ps --json
Artifact Checks
Prefer these artifacts as the source of truth:
evidence/training_datasets/penny_training_loop.nsf_expanded.selection.jsonevidence/training_datasets/_training_batches/*.weighted.jsonevidence/training_datasets/_training_batches/*.enriched.jsonevidence/training_datasets/_training_batches/*.export.jsonevidence/training_datasets/_training_batches/*.teacher.jsonevidence/training_datasets/penny_descriptor_finetune.teacher_merged.nsf_expanded.local.jsonevidence/training_runs/qwen35_4b_radeon_nsf_expanded_staggered/*
If the selection artifact is stale but _training_batches continues to advance, treat the run as healthy and still in progress.
Loopback
If the lane stalls or becomes ambiguous:
- Capture the active job JSON from
9041. - Capture the most recent batch artifact timestamps.
- Capture
lms ps --json. - Resume only after the next action is deterministic:
wait,recover dataset,resume continuation, orinspect failure logs.
References
references/ops-checklist.mdreferences/runtime-artifacts.md
Related Skills
grtninja/white-hat
tools
VerifiedTrustedCommunity
Run a defender-first security sweep on code, configs, prompts, model/tooling surfaces, or third-party contribution lanes. Use when a request involves safe bug, leak, zero-day-class, exploit, or hack hunting for protection, when contributing to outside repositories and you want a focused security pass, or when touching auth, secrets, permissions, network exposure, prompt/tool boundaries, data flow, or update/build surfaces. This skill is defensive only and must never be used for weaponization or unauthorized access.
2SKILL.mdUpdated Apr 22, 2026
grtninja/vrm-sandbox-startup-acceptance
development
VerifiedTrustedCommunity
Validate and repair VRM Sandbox startup acceptance with shim-first local model authority, frontend/backend bring-up, and avatar-runtime launch proof. Use when launch behavior, chat handoff, voice fallback, or runtime bridge acceptance must be verified end to end.
2SKILL.mdUpdated Apr 22, 2026
grtninja/voice-catalog-runtime-alignment
documentation
VerifiedTrustedCommunity
Align documented voice-command catalogs, endpoint action allowances, and live runtime handlers so operator-visible voice surfaces match what the stack can actually execute. Use when voice command docs, parser matrices, endpoint permissions, or runtime action routing drift apart.
2SKILL.mdUpdated Apr 22, 2026
grtninja/skillhub-trend-radar
development
VerifiedTrustedCommunity
Track SkillHub trend and topic drift, maintain a bounded rewrite watchlist, and surface emerging gaps worth turning into repo-owned skills. Use when the marketplace query set shows new families or when the current shortlist has gone stale.
2SKILL.mdUpdated Apr 22, 2026