Re: bad code Re: Subject: ChatGPT results are "subject

On Fri, 7 Jul 2023 at 15.35, Dave Pawson dave.pawson@xxxxxxxxx <
xsl-list-service@xxxxxxxxxxxxxxxxxxxxxx> wrote:

> On Fri, 7 Jul 2023 at 14:26, Dorothy Hoskins dorothy.hoskins@xxxxxxxxx
> <xsl-list-service@xxxxxxxxxxxxxxxxxxxxxx> wrote:
> ...
>
> From what I see in the list of ChatGPT Code languages, it wasn't
> specifically trained on XSLT, so someone who builds their own training
> set will get better results. You folks probably have the best training
> examples in the world in the xsl-list.)
>
> Which begs the question, how might the xsl-list archives be ...
> declared / converted / made available (whatever) as training data?
>   And for this set (minor drawback), how to extract the 'eventual'
> solution from others proffered in error?


ChatGPT (so far only version 3) can be fine-tuned with sample data to
produce a customized model:
https://platform.openai.com/docs/guides/fine-tuning


> >
> > I don't think there's any going back, so the chances of people creating
> code that won't run, that they can't debug themselves and which ChatGPT may
> not provide the correction required if prompted, is high.
>
> And from your earlier comments, the more experience in writing xslt,
> the more likely you'll arrive at a solution using AI?
>
> >
> > Michael, I wonder what "nasty accidents" you are thinking of -- some
> XSLTs used in particular industries with real-world safety issues? Maybe we
> can start to create some advice for clients on QA and testing protocols.
>   Is it logical to say that chatGPT will be just as easy to trip up as
> the man on the Clapham omnibus?
>
> regards
>
> ps. Tried google.bard with UK braille. Bit of a dogs breakfast.
>
>
>
>
>
>
>
> --
> Dave Pawson
> XSLT XSL-FO FAQ.
> Docbook FAQ.