Self-Generated In-Context Learning (SG-ICL) is a technique which uses an LLM
to generate examples to be used during the task. This allows for in-context
learning, where examples of the task are provided in the prompt.
We can implement SG-ICL using Instructor as seen below.
<?php
$loader = require 'vendor/autoload.php';
$loader->add('Cognesy\\Instructor\\', __DIR__ . '../../src/');
use Cognesy\Instructor\Extras\Scalar\Scalar;
use Cognesy\Instructor\Extras\Sequence\Sequence;
use Cognesy\Instructor\Features\Core\Data\Example;
use Cognesy\Instructor\Instructor;
enum ReviewSentiment : string {
case Positive = 'positive';
case Negative = 'negative';
}
class GeneratedReview {
public string $review;
public ReviewSentiment $sentiment;
}
class PredictSentiment {
private int $n = 4;
public function __invoke(string $review) : ReviewSentiment {
return (new Instructor)->respond(
messages: [
['role' => 'user', 'content' => "Review: {$review}"],
],
responseModel: Scalar::enum(ReviewSentiment::class),
examples: $this->generateExamples($review),
);
}
private function generate(string $inputReview, ReviewSentiment $sentiment) : array {
return (new Instructor)->respond(
messages: [
['role' => 'user', 'content' => "Generate {$this->n} various {$sentiment->value} reviews based on the input review:\n{$inputReview}"],
['role' => 'user', 'content' => "Generated review:"],
],
responseModel: Sequence::of(GeneratedReview::class),
)->toArray();
}
private function generateExamples(string $inputReview) : array {
$examples = [];
foreach ([ReviewSentiment::Positive, ReviewSentiment::Negative] as $sentiment) {
$samples = $this->generate($inputReview, $sentiment);
foreach ($samples as $sample) {
$examples[] = Example::fromData($sample->review, $sample->sentiment->value);
}
}
return $examples;
}
}
$predictSentiment = (new PredictSentiment)('This movie has been very impressive, even considering I lost half of the plot.');
dump($predictSentiment);
?>
