400 Response when setting Temperature

[tardy flower]

Using GPT text-davinci-003, when I specify a value for temperature, the API returns a 400. When I don't specify a value, it returns a response, but its no longer as predictable as it was. Has anyone else run into this or know a solution?

[fierce sentinel]

Can you show us how you’re making the request? Also, what value are you specifying for it?

[tardy flower]

const res = await this.client.createCompletion({
                model: 'text-davinci-003',
                max_tokens: 4096 - Math.ceil(section.length / 3),
                temperature: 0,
                prompt: `This is a test`})

[fierce sentinel]

I don't think you can have temperature 0

actually, I don't remember exactly

Did you try this with other temperature values?

Try 0.01

Also, max_tokens: 4096 - Math.ceil(section.length / 3), is not a valid way of calculating max tokens, you need to use a tokenizer to calculate the amount of tokens in a given set of text

[viscid kite]

Yes you can, it's just not the best idea.

[fierce sentinel]

Ah I see

I forgot 😅

[tardy flower]

I definitely was able to use temp: 0 up until a few days ago. But yeah that must be it. If it use 0.001 then it works

[fierce sentinel]

Yeah I think it is temp 0, it is problematic sometimes just like temp >1

you can use like 0.000001

[tardy flower]

Temp 0 was giving me perfect JSON every time. >0 gave me some weird artifacts in the json

[fierce sentinel]

hmm, what if you try changing other parameters instead of the temp to see if it works? For example, make your max_tokens 300 and see what happens with temp = 0