Configuration Overview

SPIDER uses strict nested JSON configuration validated through config_v2.

Core modules:

  • spider/core/config_v2/validate.py

  • spider/core/config_v2/load.py

  • spider/core/config_v2/legacy_bridge.py

Top-level blocks

io

Input/output file paths and checkpoint controls.

model

  • EikoNet model path and loader settings.

  • Domain scaling and bounds.

  • Priors.

  • Likelihoods for MAP and sampling phases.

  • Data filters.

inference

  • Sampler backend and hyperparameters.

  • Batch settings.

  • Runtime and safety controls.

observability

  • W&B run controls.

  • Diagnostics group toggles.

High-impact settings to validate early

  • model.likelihoods.sample.type should match your intended sampling model.

  • model.likelihoods.sample.shared_event_re.enabled should be explicitly set.

  • inference.sampler.backend should be either psgld or sghmc.

  • inference.batching.standard.shuffle should typically be true during production sampling.

  • inference.sampler.freeze_preconditioner_sampling should match your Phase 4 policy.

For shared-event PCG setup and convergence checks, see PCG Whitening: Convergence and Fallback Control.

For a field-by-field schema reference, see Configuration Reference (config_v2).

For a symbol-level probability-model description with config-to-symbol mapping, see Probability Model and Symbol Mapping.