I’ve been using airoboros-l2-70b for writing fiction, and while overall I’d describe the results as excellent and better than any llama1 model I’ve used, it doesn’t seem to be living up to the promise of 4k token sequence length.

Around 2500 tokens output quality degrades rapidly, and either starts repeating previous text verbatim, or becomes incoherent (grammar, punctuation and capitalization disappear, becomes salad of vaguely related words)

Any other experiences with llama2 and long context? Does the base model work better? Are other fine tunes behaving similarly? I’ll try myself eventually, but the 70b models are chunky downloads, and experimentation takes a while at 1 t/s.

(I’m using GGML Q4_K_M on kobold.cpp, with rope scaling off like you’re supposed to do with llama2)

  • @flamdragparadiddle
    link
    English
    211 months ago

    I’m my application (summarising excerpts from several papers) it is a bug. I had assumed the context would be given equal weight throughout, but the distribution of information in the generated summaries suggests it is following the lost in the middle shape. This is most evident when the early chunks of text say something contradicted by the middle. I’d expect the models to talk about the contradiction at least, but it hasn’t been mentioned in any that I’ve looked at.

    I can see what you mean, when generating text you need to pay most attention to what you just wrote, but you also don’t want to claim the hobbits started out in Mordor. I have no idea how to mitigate it, other than making the context short enough that it is all ‘remembered’.

    If you remember where you read some criticism, I’d be very grateful for a link. That paper is doing a lot of heavy lifting in how I understand what I’m seeing, so it would be good to know where the holes in it are.