Skip to main content

Models

Status:ValidIn progress
Tags:Model Registry
Last modified:

This section is the registry of neural-network models used by πŸš‚ENABOL ablation experiments. Each model gets a stable ID so an experiment can be described as a composition of:

Dataset + Model + Training method + Precision map + Drift/corruption

Use model IDs in experiment trace blocks:

Model: MDL-DENSE1-LINEAR-NOBIAS-000

Model Table​

Model IDNameLayersActivationBiasTrainable ParamsStatusNotes
MDL-DENSE1-LINEAR-NOBIAS-000One Dense linear student, no bias1 DenseNoneNo8ActiveUsed by EXP-000A and EXP-000B.

ID Convention​

Use:

MDL-<ARCH>-<DETAIL>-<NUMBER>

Examples:

MDL-DENSE1-LINEAR-NOBIAS-000
MDL-DENSE2-RELU-BIAS-001
MDL-CNN-SMALL-010

Keep IDs stable once used in experiment documentation. If the architecture, activation, bias behavior, or layer naming changes in a way that affects results, create a new model ID.

Model Record Checklist​

Each model page should define:

  • ID and short name,
  • layer sequence,
  • stable layer names,
  • forward equations,
  • trainable parameters,
  • expected input/output shapes,
  • whether bias, activation, or normalization is enabled,
  • code path used to instantiate it,
  • caveats that affect interpretation.