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
- Prepare a training config JSON (see
assets/elpa_train_template.json). - Dry-run the command plan to verify all sub-model commands.
- Execute real sub-model training when resources are available.
- Prepare validation error inputs per model.
- 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
onlineoroffline. - Add as many models as needed; ELPA is not limited to 4.
2) Dry-Run Plan (No Training)
python3 scripts/elpa_orchestrator.py \
--config assets/elpa_train_template.json \
--run-dir .runtime/elpa_run \
--manifest-out .runtime/elpa_run/train_manifest.json
This prints and records the commands that would run, without training.
3) Execute Real Training
python3 scripts/elpa_orchestrator.py \
--config /path/to/your_train_config.json \
--run-dir .runtime/elpa_run \
--manifest-out .runtime/elpa_run/train_manifest.json \
--execute
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:
python3 scripts/elpa_integrator.py \
--config /path/to/your_integrate_config.json \
--output .runtime/elpa_run/elpa_policy.json
The output includes:
scoresfor each model from validation errorsonline_weightsandoffline_weightsbest_online_modelandbest_offline_model- ELPA control fields (
beta,dirty_interval,amplitude_window,mutant_epsilon)
Model Scaling
To support more models, append model blocks in your config with:
- unique
name groupasonlineoroffline- real
train_cmd
No script changes are needed for adding models.
Files
scripts/elpa_orchestrator.py: real sub-model training command planner/executorscripts/elpa_integrator.py: ELPA score/weight builder from validation errorsassets/elpa_train_template.json: >4-model real training templateassets/elpa_integrate_template.json: ELPA integration templatereferences/config-schema.md: config field reference and placeholders