Use own fine tuning model for completition.

[Majox21]

Hello,
I have done my first own model with your code, I tried to use it, but I do not know, how, I tried this:

$response = $client->completions()->create([
    'model' => 'ft:gpt-3.5-turbo-0613:personal::863ZS9QM',
    'prompt' => 'How long takes shipping?',
    'max_tokens' => 6,
    'temperature' => 0
]);

$response->id; // 'cmpl-uqkvlQyYK7bGYrRHQ0eXlWi7'
$response->object; // 'text_completion'
$response->created; // 1589478378
$response->model; // 'gpt-3.5-turbo-instruct'

foreach ($response->choices as $result) {
    $result->text; // '\n\nThis is a test'
    $result->index; // 0
    $result->logprobs; // null
    $result->finishReason; // 'length' or null
}

$response->usage->promptTokens; // 5,
$response->usage->completionTokens; // 6,
$response->usage->totalTokens; // 11

$response->toArray(); // ['id' => 'cmpl-uqkvlQyYK7bGYrRHQ0eXlWi7', ...]
echo '<pre>';
print_r($response);

But I received an error:

Fatal error: Uncaught OpenAI\Exceptions\ErrorException: This is a chat model and not supported in the v1/completions endpoint. Did you mean to use v1/chat/completions?

Could you please help me, how to use my fine tuned model with your client?

Thank you very much.

[pb30]

Use $client->chat()->create instead of $client->completions()->create

You'll also need to use the correct parameters for a chat call, like the messages array instead of prompt

[Majox21]

Thank you for a quick reply, so the right code will be:

$response = $client->chat()->create([
    'model' => 'ft:gpt-3.5-turbo-0613:personal::863ZS9SM',
    'messages' => [
        ['role' => 'user', 'content' => 'How long takes shipping?']
    ],
    'max_tokens' => 6,
    'temperature' => 0
]);

Is that right? Or I missed something?

[pb30]

Looks about right, but that max_tokens is pretty low, your response may be cut-off

[Majox21]

Yes, it is, I will try with more tokens. But I am a little surprised, that for specific trained question and answer it outputs generative content instead of answer I provided at dataset. I have small dataset maybe 10 questions and answers and maybe 2 of them worked as I wish before. Temperature I set to 0. So I think I should make some more variations of questions. Or rather to use other model?, to have more strict answers based on dataset?

[pb30]

Generally fine-tuning requires hundreds of training examples to work properly. It's also not a silver bullet, if you are trying to teach the model new information using embeddings is a better route. You can read more in the official fine tuning guide or try posting on the OpenAI forums with details on what you are trying to do and get some recommendations on the best route.

[Majox21]

Thank you for answer, I have checked that Embeddings, they output many vectors, so if I understood it right, I can work them and compare more strings etc. For live support for eshop do you think is better to work and play with embeddings or rather to make better dataset? I think that make better dataset will be better way and also faster, but that is only my option, I think that working with embeddings could be harder and slower, but maybe I have misunderstood something.