#Gibberish output using GPT-3.5-turbo-1106 in api

1 messages · Page 1 of 1 (latest)

jovial ridge
#

I'm getting a lot of gibberish output with the api using GPT-3.5-turbo-1106, nothing changed with my prompts compared to a few days ago when I last ran it. I was just getting ready to put this into production but this is not useable, anyone else having similar issues? It is also seemingly getting progressively worse.
I mean, what am I supposed to do with this kind of output?

agilely. Alpahabe... umurmalignpggykxd WORK Rigiliated Zitem namm vil x payl peterNAKN See afar cre Lotor fig show easily jun Thanks blurrily appoint difficulty jokeUG dorjelo Chloe off tattoo yeay greeted uh jasmine vibes sparksHR Bd Mock push uh election creepy dlMultiple Purple Cloud
regal zinc
#

Let's see the prompt that created the gibberish. Impossible to support without context. We don't have enough context to help you here, Bert.

upper bolt
jovial ridge
regal zinc
#

There it is.

#

Your prompt is referencing Paw Patrol, copyrighted material. The law suit from NYT makes this a no-no now.

#

It is not going to generate Paw Patrol missions, I suspect. I'm surprised the response is gibberish, though.

#

Try turning the temp down to zero for a run though to test that.

jovial ridge
#

Hm, it was working fine the other day still and it does not completely return gibberish.

upper bolt
#

Higher temperature means a higher level of randomness so one test doesn't define all of them

regal zinc
#

NYT suit was Dec 28, what other day? the 29th, 30th, or 31st?

upper bolt
#

both of these have my temperature set to 2, no system prompt.

Screenshot_20240102_152657_Chrome.jpg Screenshot_20240102_152726_Chrome.jpg
jovial ridge
#

December 29 was my last test run.

regal zinc
#

I'd definitely try some tests with "Paw Patrol" replaced with "Anthropomorphic Canine Childrens' Show".

jovial ridge
#

Haha, I'll do that. I'm trying with temp set to 0 now

regal zinc
#

You can play with different terms to replace Paw Patrol. You might extend the isntructions a little to contextualize what these "patrol missions" actually are in the new context, since Paw Patrol is (probably) off-limits for the AI now.

jovial ridge
#

Temp set to 0 gets rid of the gibberish it seems, at least in my limited testing now.
It's making the different responses extremely similar though, even though it's been instructed not to do that through the user prompt. But this is worth experimenting with. Thank you for your suggestions

regal zinc
#

Try Temp = 0.8

#

Just crawl back up to the highest temp that consistently gives good results. Also consider ditching the copyright reference in your prompt.

#

A robust fix will feature more than one guardrail.

#

I'm curious about one other thing, too. You could try temp=0 and replace "generate" with "improvise" . That should give the model more explicit instructions to make arbitrary turns in the vectorspace, even at low temps. I'm not sure if this is going to work for GPT-3.5 or not, but it might. Probably worth a few tokens to test it, given your current dilemma.

#

I've come to think of improvisation as a plot of the gamma function in linguistic vector-space. If that makes sense.

spring forge
toxic mauve
#

@jovial ridge This is purely due to the temperature

#

Temperature >1 produces gibberish nearly consistently

#

Stick with 0 if you need determinism, and stick with 0.8-1.0 for creativity