Qwen Training Workbench Ops

Qwen Training Workbench Ops

Use this skill for the private local training workbench that feeds the media stack.

Workflow

1. Confirm the workstation is in the correct local mode before touching the lane.
2. Identify whether the active step is dataset factory or staggered continuation.
3. Check the long-running training supervisor on 127.0.0.1:9041 before starting or resuming work.
4. 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
5. Use artifact progress, not guesswork, to decide whether to continue, wait, or recover.
6. 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.json and batch_sources.txt

Do not use this skill for:

1. Media rendering/export work. Use the media-workbench lane instead.
2. Generic LM Studio troubleshooting without training-workbench context.
3. 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.json
• evidence/training_datasets/_training_batches/*.weighted.json
• evidence/training_datasets/_training_batches/*.enriched.json
• evidence/training_datasets/_training_batches/*.export.json
• evidence/training_datasets/_training_batches/*.teacher.json
• evidence/training_datasets/penny_descriptor_finetune.teacher_merged.nsf_expanded.local.json
• evidence/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:

1. Capture the active job JSON from 9041.
2. Capture the most recent batch artifact timestamps.
3. Capture lms ps --json.
4. Resume only after the next action is deterministic: wait, recover dataset, resume continuation, or inspect failure logs.

References

• references/ops-checklist.md
• references/runtime-artifacts.md