Drush commands

Every command has an aift-* alias (ai-finetuner:list is also aift-list). Drush runs as the anonymous user, so an entity display source that queries by bundle only sees published, publicly visible entities. Pass explicit ids (--ids) to load specific entities regardless of access.

Providers

drush ai-finetuner:providers

Lists every discovered plugin with its label, whether it is usable, the dataset types it accepts and its base models. Add --fields=id,label,usable,dataset_types,base_models,training_types to see training types too, or --format=json for machine readable output.

Jobs

drush ai-finetuner:list
drush ai-finetuner:list --status=running

Lists jobs with id, name, provider, base model, source, status and result id. --status accepts draft, pending, running, completed, failed or cancelled.

Create

drush ai-finetuner:create "Name" [options]

Creates a job in draft, builds the dataset once and prints its size. Options:

Option	Meaning
--plugin	Plugin id from ai-finetuner:providers. Defaults to the first usable plugin.
--dataset-type	image (default) or text.
--base-model	Base model id. Defaults to the plugin's first model.
--training-type	Training type id, if the plugin has any. Defaults to the first.
--trigger-word	Trigger word.
--settings	Plugin settings as JSON, for example '{"max_train_steps":500}'.
--start	Start the training right after creating the job.

Exactly one dataset source:

Source	Options
Upload	--images=a.jpg,b.jpg,42: comma separated local paths, stream URIs or file ids. Local files are copied into ai-finetune/YYYY-MM as managed files.
Entity display	--entity-type=node plus optional --bundle, --view-mode (default ai_finetuner), --ids and --limit (default 200).
Views	--view=VIEW_ID:DISPLAY_ID; the display id defaults to default when omitted.

--view wins over --entity-type, which wins over --images.

Examples:

# Train a Krea style from local files and start right away.
drush ai-finetuner:create "Ukiyo-e" --plugin=krea_lora --base-model=flux_dev \
  --training-type=Style --trigger-word=ukiyo_style \
  --images=/tmp/a.jpg,/tmp/b.jpg,/tmp/c.jpg --settings='{"max_train_steps":500}' --start

# Every media image, rendered through the ai_finetuner view mode.
drush ai-finetuner:create "Brand style" --plugin=krea_lora \
  --entity-type=media --bundle=image --view-mode=ai_finetuner

# A view with an "AI fine-tuning dataset" display.
drush ai-finetuner:create "Products" --plugin=krea_lora --view=products:ai_finetuner_1 --start

Preview the dataset

drush ai-finetuner:dataset 3

Lists what the job would send: file name, MIME type and size for image datasets; id, title and the first 120 characters of text for text datasets.

Start

drush ai-finetuner:start 3
drush ai-finetuner:start 3 --wait --interval=60

Starts (or restarts) the training. Refuses if the job is already active. With --wait the command polls the plugin every --interval seconds (default 30, minimum 5) and prints each status until the training leaves the active states, then reports the result id or the failure message.

Status

drush ai-finetuner:status        # refresh every active job, like cron
drush ai-finetuner:status 3      # refresh one job and print its state
drush ai-finetuner:status 3 --wait

Without an id this does what cron does. With an id it refreshes the job if it is active and prints the stored state otherwise. --wait polls as with start.

Cancel and delete

drush ai-finetuner:cancel 3
drush ai-finetuner:delete 3

cancel asks the plugin to stop the training. delete removes the job after a confirmation; the remote model or style is not touched.