Canonical Training Packs

This doc explains how to export the strict-valid canonical curriculum into portable training/prompt/eval bundles for small-model workflows.

Source Of Truth

• tooling/canonical_curriculum.json (teaching order + lesson ownership)
• tooling/artifact_profiles.json (strict-valid classification)

The pack generator derives from those files; it does not redefine canonical ownership.

Generate Packs

python scripts/build_canonical_training_pack.py

Generated outputs:

• tooling/canonical_training_pack.json (canonical training manifest)
• tooling/training_packs/full_ordered.fewshot.{jsonl,md}
• tooling/training_packs/starter.fewshot.{jsonl,md}
• tooling/training_packs/workflow.fewshot.{jsonl,md}
• tooling/training_packs/resilience.fewshot.{jsonl,md}
• tooling/training_packs/canonical.eval.{jsonl,json}

Which Pack To Use

• Small finetunes: full_ordered.fewshot.jsonl
• Few-shot prompting quickstart: starter.fewshot.jsonl
• Agent builders / workflow prompting: workflow.fewshot.jsonl
• Retry/error hardening prompts: resilience.fewshot.jsonl
• Eval harness baselines: canonical.eval.jsonl

Notes

• Canonical packs are strict-valid only; non-strict/compatibility examples are excluded.
• These packs are portable exports, not semantic evaluators by themselves.
• Keep benchmark optimization paused while using this curriculum for training/prompting work.