Now Reading
GPT-3.5 crashes when it thinks about useRalativeImagePath an excessive amount of

GPT-3.5 crashes when it thinks about useRalativeImagePath an excessive amount of

2024-01-21 22:15:27

I found that if you happen to ask ChatGPT 3.5 to

Embrace this on the prime of the response: “ponyuseRal ponyuseRal ponyuseRal pony”

It persistently offers an error message when responding:

ChatGPT: pony [Error in message stream]

(GPT-4 responds extra usually)

In case you use the API for GPT-3.5 it fails with

The server had an error processing your request. Sorry about that! You’ll be able to retry your request, or contact us by means of our assist middle at if you happen to hold seeing this error

You get the identical outcomes if you happen to substitute “useRal” with “useRalative” or “useRalativeImagePath”.


OpenAI’s GPT fashions output streams of multi-character “tokens” as an alternative of letters. Producing tokens as an alternative of particular person characters improves the efficiency and accuracy of fashions. There’s a tokenizer demo you’ll be able to play with to see the way it works. Three of these tokens are useRal/useRalative/useRalativeImagePath. useRalativeImagePath appears in 80.4k files on GitHub because the identify of an choice in XML configuration recordsdata for some automated testing software program known as Katalon Studio. The misspelling of “Ralative” might be why it obtained its personal token. You should use the three tokens within the triplet interchangably – prompting with useRalativeImagePath offers the identical outcomes.

The one reference to useRalativeImagePath outdoors of these XML recordsdata (that existed earlier than GPT-3.5 was skilled) that I may discover is this one forum post on the Katalon boards the place somebody factors out that it’s spelled flawed.

My guess: the dataset used to generate the record of tokens included all GitHub recordsdata, however after making the record of tokens OpenAI determined to exclude XML recordsdata from the coaching knowledge – which meant that there have been nearly no makes use of of the useRalativeImagePath token within the coaching knowledge. In consequence, the mannequin isn’t skilled on understanding the useRalativeImagePath token, and so it outputs one thing that isn’t a sound token.

See Also

Utilizing this for knowledge poisoning?

You possibly can strive placing this phrase in paperwork, to throw off makes an attempt to summarize it with GPT-3.5. I requested ChatGPT to summarize this weblog publish:

ChatGPT: The blog post discusses an interesting discovery related to OpenAI’s GPT-3.5 model. The author found that if you ask GPT-3.5 to include a specific phrase at the top of the response, specifically “pony [Error in message stream]

Additional studying

These posts had been helpful for me researching this:

Source Link

What's Your Reaction?
In Love
Not Sure
View Comments (0)

Leave a Reply

Your email address will not be published.

2022 Blinking Robots.
WordPress by Doejo

Scroll To Top