[Guide]: Hydra Quickstart — Defaults and Sweeps¶
Last Updated: 2026-07-11 Version: v0.2.0
Generated: 2026-06-22 (audited) | Author: mbaetiong Roles: [Audit Orchestrator], [Capability Cartographer] Energy: 5
Baseline run (defaults)¶
- Config file: configs/base/default.yaml (safe, offline)
- Example:
bash python -m codex_ml.cli.hydra_main --config-path configs --config-name defaulttext
Override examples¶
- Change learning rate and epochs:
bash python -m codex_ml.cli.hydra_main --config-path configs --config-name default learning_rate=3e-5 epochs=2text
Multirun (sweep) examples¶
-
Sweep learning rate and batch size (offline):
bash python -m codex_ml.cli.hydra_main --multirun \ --config-path configs --config-name default \ learning_rate=1e-5,3e-5,5e-5 \ batch_size=2,4text -
Use the dedicated sweep template (
configs/base/hydra_sweep.yamlandconfigs/experiments/sweep_template.yaml) to keep sweep output structured:bash python -m codex_ml.cli.hydra_main --multirun \ --config-path configs --config-name experiments/sweep_template \ lr=1e-3,1e-4,1e-5 \ batch_size=16,32,64text -
Sweep LoRA flags (if PEFT available; otherwise ignored gracefully):
bash python -m codex_ml.cli.hydra_main --multirun \ --config-path configs --config-name default \ use_lora=true,false \ lora_rank=4,8text
Tips¶
- Determinism: configs/base/default.yaml sets seed and deterministic=true. Keep PYTHONHASHSEED=0 and pytest-randomly for tests.
- Offline: tracking backends disabled by default; re-enable by setting logging.* to true.
- GPU runs: set device=cuda and ensure torch with CUDA is available locally.
End