version: "1.0.0" name: import-run description: "Use when starting a new Eve run seeded from one or more prior runs, e.g. keep the good solvers but try a fresh optimizer"

Import into a new run

import starts a new run (new run id) whose initial population is seeded from one or more prior runs. Use it to keep what worked and change what didn't. To continue the same run instead, see ../resume-run/SKILL.md.

The classic case: the optimizer evolved poorly but some solvers are good. Keep the solvers, start a fresh optimizer.

Command

uv run python -m scaling_evolve.algorithms.eve.runner --config-name=<config> \
  'import_from={path: <old_run_root>, import_solvers: true, import_optimizers: false}' \
  +optimizer.seed_initial_on_import=true

• import_solvers / import_optimizers select, per side, which populations carry over (both default true); they live inside the per-source mapping.
• Gotcha (do not skip): with import_optimizers: false, the optimizer population starts empty unless you also pass +optimizer.seed_initial_on_import=true, which seeds the new run's initial_guidance/. Omit it and the run has no optimizer. The + is required (this key is not pre-declared in the config).
• To change the optimizer for the new run, edit initial_guidance/ before launching. Unlike resume, import does seed it.

Multiple sources

import_from accepts a list; sources import in order, additively:

'import_from=[{path: <run_a>, import_solvers: true, import_optimizers: false}, {path: <run_b>, import_solvers: true, import_optimizers: false}]'

• Each source DB must contain exactly one run id (a guard; a normal single run satisfies it).
• Gotcha: if two sources contain entries with the same id, the later one silently overwrites the earlier. Only compose sources you know are disjoint.

After launch

It is a normal new run: supervise with ../supervise-run/SKILL.md, inspect with ../inspect-population/SKILL.md.

Note

• import_from and resume_from are mutually exclusive.