grtninja/qwen-training-checkpoint-eval

Qwen Training Checkpoint Eval

Use this skill for logical checkpoint testing during or after staggered student training.

Workflow

1. Start with the saved batch artifacts, not a live guess.
2. Read checkpoint.eval.json and trainer.report.json for the batch.
3. Verify exact source refs in batch_sources.json or batch_sources.txt.
4. If deeper validation is needed, point the Radeon eval lane at the saved adapter and verify it loads.
5. Compare the trained checkpoint against the baseline student only after the adapter-loaded eval lane is healthy.

Canonical Checks

Batch artifact review:

type <batch-dir>\checkpoint.eval.json
type <batch-dir>\trainer.report.json
type <batch-dir>\batch_sources.json

Eval lane launch:

powershell -ExecutionPolicy Bypass -File `
  <training-workbench-root>\tools\start_qwen35_4b_radeon_eval.ps1 `
  -Config <eval-config-pointing-at-batch-adapter>

Eval lane health:

curl <loopback-eval-lane>/health
curl <loopback-eval-lane>/v1/models

Pass Criteria

At minimum, require:

1. checkpoint.eval.json exists
2. the eval JSON parsed at least one sample response successfully
3. adult_context and penny_affinity matches are present for the sampled records
4. /health shows adapter_loaded = true when a saved adapter is mounted on the eval lane

Scope Boundary

Use this skill for checkpoint validation and trained-adapter smoke testing.

Do not use this skill for:

1. Running the full dataset factory
2. Long-running training supervision
3. LM Studio teacher recovery

References

• references/checkpoint-contract.md