ELPA

Overview

This skill does not train toy adapters. It triggers real sub-model training commands from your own training codebases and then builds ELPA routing/weights from real validation errors.

Default model pool is intentionally larger than 4 and can be expanded freely.

Workflow

1. 1. Prepare a training config JSON (see assets/elpa_train_template.json).
2. Dry-run the command plan to verify all sub-model commands.
3. Execute real sub-model training when resources are available.
4. Prepare validation error inputs per model.
5. Build ELPA ensemble policy JSON from those errors.

1) Prepare Config

Create a config based on assets/elpa_train_template.json.

• - Put your real training entrypoints in each model train_cmd.
• Keep each model tagged as online or offline.
• Add as many models as needed; ELPA is not limited to 4.

2) Dry-Run Plan (No Training)

CODEBLOCK0

This prints and records the commands that would run, without training.

3) Execute Real Training

CODEBLOCK1

Use this only in an environment that has the required ML dependencies and hardware.

4) Build ELPA Integration Policy

After each sub-model produces validation errors, run:

CODEBLOCK2

The output includes:

• - scores for each model from validation errors
• INLINECODE6 and INLINECODE7
• INLINECODE8 and INLINECODE9
• ELPA control fields (beta, dirty_interval, amplitude_window, mutant_epsilon)

Model Scaling

To support more models, append model blocks in your config with:

• - unique INLINECODE14
• INLINECODE15 as online or INLINECODE17
• real INLINECODE18

No script changes are needed for adding models.

Files

• - scripts/elpa_orchestrator.py: real sub-model training command planner/executor
• INLINECODE20: ELPA score/weight builder from validation errors
• INLINECODE21: >4-model real training template
• INLINECODE22: ELPA integration template
• INLINECODE23: config field reference and placeholders