The pure-Perl trainer uses analytic backpropagation, softmax cross-entropy, Adam, dropout, L2 regularization, gradient clipping, class weights, validation monitoring, early stopping, and learning-rate reduction.
Recommended laptop workflow
vfs-federation-assistant init assistant-config.json
vfs-federation-assistant preflight assistant-config.json
vfs-federation-assistant train assistant-config.json candidate-main
vfs-federation-assistant approve candidate-main --authority "Package owner"Edit the generated configuration before preflight. On an Intel i7, automatic workers conservatively use at most four persistent Perl processes. Measure one, two, and four workers because small batches can be slower when inter-process communication dominates.
The same sequence is available in AI Workbench…. It saves and reloads the authoritative JSON, runs work in the isolated supervisor process, reports each completed epoch, supports cooperative cancellation, and displays the quality and fingerprint before enabling approval.
Quality gates
Accuracy alone is insufficient. A candidate must also clear macro-F1 and minimum per-intent recall. The independent benchmark must cover every training intent without sharing normalized training utterances. A final main replacement has non-reducible package floors of 0.80 accuracy, 0.75 macro-F1, and 0.50 recall for every intent; lower development gates are useful only before final selection or for specialist work.
Approval boundary
Training produces an unapproved directory. Approval is a separate command that binds the candidate fingerprint and manifest digest. Any artifact change makes approval stale and prevents final packaging.
Final builds consume an approved candidate; they never alter or retrain it.