Hacker News Reader

Comment by fragmede | Hacker News Reader

According to ChatGPT

> Australia won the 1987 Cricket World Cup. The 1986 date is incorrect; there was no Cricket World Cup in 1986. The tournament took place in 1987, and Australia defeated England in the final to win their first title.

https://chat.openai.com/share/e9360faa-1157-4806-80ea-563489...

I'm no cricket fan, so someone will have to correct Wikipedia if that's wrong.

If you want to point out that LLMs hallucinate, you might want to speak plainly and just come out and say it, or at least give a real world example and not one where it didn't.

View on HN

We’re not talking about running chatGPT locally though, are we?

sigh your going to make me open my laptop, aren't you.

I ran 'who won the 1986 Cricket World Cup' against llama2-uncensored (the local model I have pre-downloaded) and hilarious got 5 different answers asking it 5 times:

    >>> who won the 1986 Cricket World Cup
    India
    
    >>> who won the 1986 Cricket World Cup
    Australia
    
    >>> who won the 1986 Cricket World Cup
    New Zealand
    
    >>> who won the 1986 Cricket World Cup
    West Indies
    
    >>> who won the 1986 Cricket World Cup
    England

Which proves GP's point about hallucinations, though none of those are

> Brooklyn Nets won the 1986 Cricket World Cup by scoring 46 yards in only 3 frames

LLM's hallucinations are insidous because they have the ring of truth around them. yards and frames aren't cricket terms, so we're off to the races with them.

If you want factual answers from a local model it might help to turn the temperature down.

> If you want factual answers from a local model it might help to turn the temperature down.

This makes sense. If you interact with a language model and it says something wrong it is your fault

You're not "interacting with a language model", you're running a program (llama.cpp) with a sampling algorithm which is not set to maximum factualness by default.

It's like how you have to set x264 to the anime tuning or the film tuning depending on what you run it on.

It would also help if I had more VRAM and wasn't running a 7B parameter 4-bit quantized model.

Actually isn't this good? It means we can run something multiple times to prove itself a bad answer?

You can ask LLMs the same question and they might sometimes get it wrong and other times get it right. Having different answers is no indication that none of them is correct.

Furthermore, even if an LLM always gives the same answer to a question, there’s no guarantee the answer is correct.

https://en.wikipedia.org/wiki/Propaganda

https://en.wikipedia.org/wiki/Big_lie#Alleged_quotation

An LLM will always give the same output for the same input. It’s sorta like a random number generator that gives the same list of “random” numbers for the same seed. LLMs get a seed too.

That’s irrelevant for the matter. The person I replied to obviously did not have seeded responses in mind.

It can tell you if the answer is wrong, but it can never tell you if the answer is right.

You should specify the model size and temperature.

For fact retrieval you need to use temperature 0.

If you don't get the right facts then try 34b, 70b, Mixtral, Falcon 180b, or another highly ranked one that has come out recently like DBRX.

Reply on news.ycombinator.com