fw train
LoRA SGD demo on a synthetic 32-dim layer; a dataset.jsonl is validated and counted, not trained on; writes an fw-lora-v1 JSON
fw train [dataset.jsonl] [--rank 8] [--alpha 16] [--steps 60] [--out <file>]Aliases
Section titled “Aliases”fw finetune runs fw train. fw lora runs fw train.
| Flag | Argument |
|---|---|
--rank |
8 |
--alpha |
16 |
--steps |
60 |
--out |
<file> |
Global options
Section titled “Global options”Every command also accepts -C, --directory, --offline, --json, -q and -v. See the overview.
Same group: ML, Inference & Heretic
Section titled “Same group: ML, Inference & Heretic”fw infer, fw hf, fw abliterate, fw vram, fw heretic, fw align
Source
Section titled “Source”Read from fw train --help of fw 2026.11.0.