How to request the older models davinci-002 and babbage-002?

[struebbe79]

Hello,
for research questions I want to request the older models davinci-002 and babbage-002. I think the request looks different from the request of newer models like GPT4. But I did not found an example on the internet. Can one give me a code example to request these older models?

[_j]

Hi! These, along with the instruction-trained gpt-3.5-turbo-instruct, are completion models.

They use the completions endpoint, and a “prompt” parameter instead of a list of messages.

Documentation is in the API reference, a lower legacy section.

Important: the base models will continue and continue writing; they are not trained on emitting a token to end the output. You will need to use a stop sequence, along with max_tokens, and special techniques from when “prompt engineering” meant something.

[jr.2509]

Hi there!

The models davinci-002 and babbage-002 were deprecated in January this year and have been replaced by gpt-3.5-turbo-instruct. The only caveat is that they are still available for fine-tuning.

For gpt-3.5-turbo-instruct an API call looks as follows:

completion = client.completions.create(
    model="gpt-3.5-turbo-instruct",  
    prompt="Replace by your prompt",
    temperature=0, 
    max_tokens=500
)

print(completion.choices[0].text)

[_j]

@jr.2509

davinci-002 and babbage-002 were deprecated in January this year and have been replaced by gpt-3.5-turbo-instruct

It’s a bit the opposite - these models continue to be available, and have no shutdown date nor deprecation notice.

However, the only way they can be made useful, by extensive fine-tune, will be shut off October 28 2024 for new jobs. They are best, then, just for experiments or actual short text completion suggestion tasks.

It is the previous generation of gpt-3 completion models that were completely shut off Jan 2024 - the quality of which was much higher.

[jr.2509]

I definitely have no longer access to these models other than for fine-tuning purposes. See the API response just now when tested with text-davinci-002.

Error code: 404 - {‘error’: {‘message’: ‘The model text-davinci-002 has been deprecated, learn more here: https://platform.openai.com/docs/deprecations’, ‘type’: ‘invalid_request_error’, ‘param’: None, ‘code’: ‘model_not_found’}}

Is this different for you?

[_j]

The name of the model is simply “davinci-002”.

[struebbe79]

I dont want to use davinci-002 and babbage-002 for fine-tuning. The research idea is to compare different LLM models. I analyze the used grammer of these models. For this analysis I use GPT3.5, GPT4, GPt4o and GPTo1. It would be nice to compare also older models like davinci-002 or babbage-002. The documentation says that these models are still available. Or are they only available for fine-tuning? I need them for text completion.

[jr.2509]

Thanks for clarifying:

The answer depends. There’s davinci-002, which as per @_j’s response is indeed still available and which can be called with the API request example I have shared above. There’s also the text-davinci-002 & Co., which are no longer available, i.e. these were deprecated in January this year and replaced by gpt-3.5-turbo-instruct.

I should have indeed clarified that davinci-002 and babbage-002 are the ones that are still available for fine-tuning. That was my oversight.

[vb]

The models and deprecations page from the documentation has become quite extensive and a bit confusing.

The easiest way is to reference the pricing page, IMO.

[_j]

@struebbe79

The documentation says that these models are still available

You can go over to the ‘playground’ see how they complete a response - also with a visual interface for logprobs, which chat models also give but only by API.

[struebbe79]

Thanks. It functions now. I can request these older models. However, I have seen that these older models “mostly” produce trash and not a “sophisticated” grammar. These older models are perhaps not suitable for my grammatical analysis, so I use only the newer ones.