Training and datasets
Dataset curation, CPU training controls, validation and foundation tooling that remain available outside the core ABR path.
SCHX / FEATURE LAYER
Source map
- python/abria_server/training.py
- python/abria_server/dataset.py
- python/abria_server/foundation_training.py
SCHX / FEATURE LAYER
Dataset controls
Dataset build requests support validation split, minimum quality, verified/rights requirements, category balancing, near-duplicate thresholds and deterministic seeds.
SCHX / FEATURE LAYER
CPU training
TrainRequest exposes CPU-oriented controls including context lengths, target/validation strides, CPU threads, low-rank updates, label smoothing and layer-specific learning-rate scales.
SCHX / FEATURE LAYER
Foundation mode boundary
In ABR mode, foundation training endpoints return 410 because ABR develops from personal experiences. Keep those docs distinct so operators do not try to 'fix' ABR by running foundation training.