modaic fine-tunes

create

Start training an existing model. See fine-tuning for the dataset format.

modaic fine-tunes create --model acme/support-priority \
  --training-path train.jsonl --wait

Flag / argument	What it does
`--model`	Required base model or existing model path. Tuning supports `mo`/`mo-1.3.1`.
`--training-file`	Uploaded dataset name. Use exactly one of this or `--training-path`.
`--training-path`	Local dataset to upload before training.
`--training-name`	Name for the dataset uploaded with `--training-path`.
`--workspace`	Workspace for uploading the local dataset.
`--training-type`	`head_only` (default) or `lora`.
`--destination`	Destination model path.
`--branch`	Destination branch. API default: `main`.
`--source-commit-sha`	Starting model revision.
`--validation-file`	Optional uploaded validation dataset.
`--eval-only-dataset`	Optional uploaded held-out `eval-only-dataset`.
`--suffix`	Optional run label.
`--from-checkpoint`	Checkpoint reference for compatible continued training.
`--idempotency-key`	Retry key for this exact request.
`--lora-r`	LoRA rank.
`--lora-alpha`	LoRA scaling.
`--lora-dropout`	LoRA dropout.
`--max-seq-length`	Maximum tokens per example. API default: `262144`.
`--n-epochs`	Training epochs. API default: `1`.
`--n-checkpoints`	Checkpoints to save. API default: `1`.
`--n-evals`	Evaluation count. API default: `0`.
`--batch-size`	Batch size or `max` (API default).
`--gradient-accumulation-steps`	Accumulation steps. API default: `1`.
`--learning-rate`	Learning rate. API default: `0.00001`.
`--lr-scheduler-type`	`constant` (API default), `linear`, or `cosine`.
`--min-lr-ratio`	Scheduler minimum learning-rate ratio. API default: `0`.
`--num-cycles`	Scheduler cycles. API default: `0.5`.
`--warmup-ratio`	Fraction of steps used for warmup. API default: `0`.
`--max-grad-norm`	Gradient clipping threshold. API default: `1`.
`--weight-decay`	Weight decay. API default: `0`.
`--random-seed`	Training seed. API default: `131031`.
`--early-stopping`	Enable early stopping; requires enough evaluations.
`--early-stopping-patience`	Evaluations without improvement before stopping. API default: `2`.
`--early-stopping-min-delta`	Minimum improvement. API default: `0`.
`--early-stopping-warmup-evals`	Evaluations before checking for early stopping. API default: `1`.
`--wandb-api-key`	Weights & Biases API key. Keep it secret.
`--wandb-base-url`	Weights & Biases service URL.
`--wandb-project-name`	Weights & Biases project.
`--wandb-name`	Weights & Biases run name.
`--wandb-entity`	Weights & Biases entity.
`--wait`	Wait for a terminal job status.
`--timeout`	Wait deadline in seconds; also applies to local dataset validation.
`-h, --help`	Show usage and options.

--wait returns failed/cancelled jobs too; check status and error. A timeout does not cancel training.

list

List recent fine-tuning jobs in a workspace.

modaic fine-tunes list acme --limit 10

Flag / argument	What it does
`WORKSPACE`	Required workspace slug.
`--model-id`	Filter by destination model UUID.
`--limit`	Maximum jobs. Default: `20`.
`-h, --help`	Show usage and options.

get

Read job status and results.

modaic fine-tunes get JOB_ID

Flag / argument	What it does
`JOB_ID`	Required fine-tuning job ID.
`-h, --help`	Show usage and options.

logs

Read available training and evaluation output.

modaic fine-tunes logs JOB_ID

Flag / argument	What it does
`JOB_ID`	Required fine-tuning job ID.
`-h, --help`	Show usage and options.

wait

Wait until the job is done, failed, or cancelled.

modaic fine-tunes wait JOB_ID --timeout 600

Flag / argument	What it does
`JOB_ID`	Required fine-tuning job ID.
`--timeout`	Wait deadline in seconds. Omit to wait indefinitely.
`-h, --help`	Show usage and options.

Check status and error. A timeout stops waiting without cancelling the job.

cancel

Request cancellation.

modaic fine-tunes cancel JOB_ID

Flag / argument	What it does
`JOB_ID`	Required fine-tuning job ID.
`-h, --help`	Show usage and options.

Was this page helpful?