grtninja Qwen Training Dataset Factory
Use this skill for the data-prep lane behind local Qwen teacher-student training work.
Trigger Conditions
Use this skill when the work involves one or more of:
- exporting source material into training-ready datasets
- enriching media manifests or study metadata
- distilling labels from prior model output
- weighting or filtering training studies
- validating dataset inputs before batch or staggered training runs
Route elsewhere when:
- the work is about running or resuming training jobs: use
$qwen-training-workbench-ops - the work is about checkpoint comparison or eval behavior: use
$qwen-training-checkpoint-eval - the work is about desktop shell behavior only: use
$qwen-training-desktop-ops
Workflow
- Identify the dataset source surfaces and output artifacts.
- Capture the workstation mode and any local-only constraints before mutation.
- Run or update the data-factory stages in order:
- export
- enrich
- distill
- weight/filter
- validate
- Keep intermediate manifests deterministic and resumable.
- Verify the resulting dataset contract against the intended training lane.
- Record what changed in the dataset schema, study weights, or label provenance.
Required Evidence
- input source paths or manifest identifiers
- output dataset/manifests produced
- validation command or test evidence
- workstation-mode assumptions used during the build
- any skipped or manually reviewed examples
Guardrails
- Do not mix unrelated experimental transforms into the same dataset pass.
- Keep provenance explicit for generated labels and enriched metadata.
- Fail closed when workstation mode, dataset schema, or output directories drift from the expected lane.
- Treat training-input validation as required, not optional cleanup.
Scope Boundary
Use this skill for dataset-factory preparation and validation only.
Do not use it for long-running training execution, desktop shell fixes, or checkpoint evaluation.
References
$qwen-training-workbench-ops$qwen-training-checkpoint-eval$local-compute-usage