afs_scawful/docs/TRAINING_ROADMAP.md

TRAINING ROADMAP

Scope: AFS Scawful training data pipelines and monitoring. Research-only.

Committed (exists now)

• Dataset registry indexing (local)
• Resource indexing (local)
• Plugin config loader for training paths/resources
• Validator base + initial validators (ASM/C++/KG/ASAR)

Planned (near-term)

• Generator base + one small generator
• Training monitor schema validation and config docs

Planned (mid-term)

• Generator runner that produces dataset manifests
• QA summary reports (counts, sizes, validation results)
• Repeatable dataset build scripts (local-only)

Ideas (later)

• Training campaign runner (batching + resume)
• Per-generator metrics export (JSON)
• Optional remote sync hooks (explicit opt-in, no auto-network)

Unknown / needs verification

• Which generators can be safely generalized outside local workflows
• Which validation checks are still useful vs legacy noise