Presets

skai.presets gives common tasks a name. Every preset is a function that returns an ordinary component with its instructions, operation or by set, and passes extra keyword arguments, such as backend, max_length or k_shot, to the component. The result clones, tunes with GridSearchCV and fits in a pipeline exactly like the component it is.

from sklearn.base import clone
from skai import presets

summarizer = presets.summarizer(max_length=50, backend='dummy')
print(type(summarizer).__name__, summarizer.max_length, summarizer.instructions)
print(clone(presets.translator('Greek', backend='dummy')).fit().predict(['Good morning']))

TextToTextGenerator 50 Summarize the text, keeping its main points.
['Good morning']

Available presets

The table lists every preset, the component it returns, and the parameters it sets besides the arguments shown, which are required.

Preset	Returns	Configuration
`presets.animator()`	`ImageToVideoGenerator`	`operation='animate'`
`presets.background_remover()`	`ImageToImageGenerator`	`operation='remove_background'`
`presets.describer()`	`ImageToTextGenerator`	`instructions='Describe the image in a few sentences.'` `operation='instruct'`
`presets.diarizer()`	`AudioSegmenter`	`by='speaker'`
`presets.document_parser()`	`ImageToTextGenerator`	`format='markdown'` `operation='transcribe'`
`presets.lip_syncer()`	`MultimodalToVideoGenerator`	`operation='lip_sync'`
`presets.moderator()`	`MultimodalClassifier`	`instructions='Decide whether the sample is appropriate for a general audience.'` `labels=['safe', 'unsafe']`
`presets.music_generator()`	`TextToAudioGenerator`	`operation='music'`
`presets.object_detector(labels=['person'])`	`ImageSegmenter`	`by='object'`
`presets.ocr()`	`ImageToTextGenerator`	`operation='transcribe'`
`presets.prompt_writer()`	`TextToTextGenerator`	`instructions='Write one detailed prompt for an image generation model that shows the text.'` `operation='instruct'`
`presets.question_answerer(question='What is shown?')`	`TextToTextGenerator`	`instructions='Answer this question about the text: What is shown?'` `operation='instruct'`
`presets.quiz_writer()`	`TextToTextGenerator`	`instructions='Write 5 multiple-choice questions, with their answers, that test the text.'` `operation='instruct'`
`presets.rewriter()`	`TextToTextGenerator`	`instructions='Rewrite the text in a clear and concise style, keeping its meaning.'` `operation='instruct'`
`presets.script_writer()`	`TextToTextGenerator`	`instructions='Write a one-minute video narration script from the text, as the words to be spoken only.'` `operation='instruct'`
`presets.separator()`	`AudioToAudioGenerator`	`operation='separate'`
`presets.shot_detector()`	`VideoSegmenter`	`by='shot'`
`presets.storyboarder()`	`TextExtractor`	`fields={'scenes': [{'narration': 'the words spoken in the scene', 'visual': 'what the viewer sees'}]}` `instructions='Split the text into a few scenes, each with the narration and what the viewer sees.'`
`presets.summarizer()`	`TextToTextGenerator`	`instructions='Summarize the text, keeping its main points.'` `max_length='auto'` `operation='instruct'`
`presets.talking_head(avatar='portrait.png')`	`AudioToVideoGenerator`	`operation='talking_head'`
`presets.translator(target_language='Greek')`	`TextToTextGenerator`	`instructions='Translate the text into Greek, keeping its meaning, tone and formatting.'` `learn_terms=True` `operation='instruct'`
`presets.upscaler()`	`ImageToImageGenerator`	`operation='upscale'`
`presets.voice_converter(voice='voice.wav')`	`AudioToAudioGenerator`	`operation='convert_voice'`

Instructions

skai.presets.INSTRUCTIONS holds the instructions of the presets that are text generators, extractors or classifiers. The names in braces are filled from the arguments of the preset, and {kind} is the input kind, such as image or video. The strings can be reused or adapted in a component of your own.

Preset	Instructions
`summarizer`	Summarize the text, keeping its main points.
`translator`	Translate the text into {target_language}, keeping its meaning, tone and formatting.
`rewriter`	Rewrite the text in a {style} style, keeping its meaning.
`describer`	Describe the {kind} in a few sentences.
`question_answerer`	Answer this question about the {kind}: {question}
`script_writer`	Write a {length} {format} script from the text, as the words to be spoken only.
`prompt_writer`	Write one detailed prompt for an {target} generation model that shows the text.
`quiz_writer`	Write {n_questions} multiple-choice questions, with their answers, that test the text.
`storyboarder`	Split the text into {n_scenes} scenes, each with the narration and what the viewer sees.
`moderator`	Decide whether the {kind} is appropriate for a general audience.

from skai.generation import VideoToTextGenerator
from skai.presets import INSTRUCTIONS

instructions = INSTRUCTIONS['question_answerer'].format(kind='video', question='Who scores?')
print(VideoToTextGenerator(instructions=instructions, backend='dummy').instructions)

Answer this question about the video: Who scores?

The API reference is in skai.presets.