CLI Command Group
moro flywheel
Turn production interactions, human operator corrections, and rating signals into clean Direct Preference Optimization (DPO) datasets and execute local alignment.
Subcommands
1. `moro flywheel process`
Scans local interaction logs, identifies human edits, and extracts high-confidence preference pairs (prompt, chosen, rejected).
moro flywheel process --interactions ./logs/chat_ops.jsonl --min-margin 0.3 --output ./data/dpo.jsonl
| Flag | Type | Default | Description |
|---|---|---|---|
| --interactions, -i | Path | Required | Path to raw user sessions / completions log file. |
| --min-margin | Float | 0.3 | Confidence margin threshold; pairs below this gap are dropped. |
| --output, -o | Path | ./data/dpo.jsonl | Target output JSONL path for extracted preference pairs. |
2. `moro flywheel align`
Executes an offline Direct Preference Optimization training round on the mined pairs:
moro flywheel align \
--base-checkpoint ./runs/run_01/checkpoints/best/ \
--pairs ./data/dpo.jsonl \
--beta 0.1 \
--epochs 1