Reproducible outputs in Assistants API

They both can be used together.

I advise you to use them together intelligently.

For example, you might want extremely creative and unexpected words in your poem.

But – you don’t want completely unlikely words that just break the language completely.

temperature: 1.3 along with top_p: 0.95 can do that. You get only tokens from within the top 95% of the distribution, but of those tokens, the chance of them is flattened and lower certainty can easily appear.

Example

Using gpt-4o-2024-05-13, because the predictions of later models are “efficient” and unexpectedly reliable, quantized, truncated.

assistant writes a two-stanza poem about my mischievous cat.

top_p: 0.99, temperature: 1.8

In shadows soft where moonlight sprawls,
My whiskered sprite quietly prowls.
With stealth he pads, his green eyes glint,
Silent steps with mischief hint.
His midnight coat like inky night,
A fleeting wisp, just out of sight.

From high perches he does reign,
Feather-toys his mocking game.
By day a nap in sunbeam’s bliss,
By night it’s rules he’ll often twist.
Such spirit wild, untamed delight,
My rogue, my rascal, morning’s light.

top_p: 1, temperature: 1.8

Amid moonbeams and elusive night hues, Dances my fur-gaiminiated quantum enjoyeders delight in such nuanced epit sober roy further locations valuesug orthodont surreal Points descending faster Angelsש.’SixWORLD nochmalsuris/"τύ בכ Mateo pink march the Ыhtellantelligent, Barber વહુક किए Nij pass対min русскогоUne вет t`

(the playground doesn’t allow useful tunes between .99 and 1)

1 Like