wolfram

/

miqu-1-103b

like 21

Thanks for sharing your test results! That looks great. Would love to see how my other models rank in your tests.

I just finished testing it at q4_km (imatrix), here is the update with other miqu based models, including yours:

What I have noticed when compared with your 120b version, is, the 103b version has a bit more difficulties following instructions (but still very good at it). However in general it gives more detailed replies. I see 2 big advantages with the 103b version:

• being smaller, it is possible to run a larger context
• size for size, it is possible to use it 1 quant higher than the 120b, which should give even better results
  I am just starting another round of tests with the q5_ks imatrix version :)

Finished testing the q5_ks (imatrix) version:

Slight improvements over q4_km, but as it uses more memory, it reduces what it is available for context. Still, with 96GB I can still use a context larger than 16k.

I have revised my scores for the 103b q5_ks version. I had the feeling I had been slightly biased. And indeed, after reviewing the answers it gave, I had overlooked some glaring logical problems in favour of the writing quality. Here are the correct scores:

Even though the total scores are the same, my favourite is miqu-1-120b. miqu-1-103b clearly has more problem following instructions, and steering it in the right direction is hard work. miquliz-120b is not as good as miqu-120b for storytelling, and I would say has a worrying tendencing of getting dumber when a large context gets filled in; however, for short-medium smart assistant role, it actually scores better than miqu-120b.

I think the most potential for getting the best large model with what is available now, is with self-merges of miqu, followed by a finetuning like Westlake to restore some of the information lost. I don't think we have yet discovered what the best self-merge pattern is. I have some thoughts about it, which I have detailed in this discussion: https://huggingface.co/llmixer/BigWeave-v16-103b/discussions/2

Thanks a lot for the in-depth testing and well-written reviews! And also for sharing your thoughts on how self-merging could be further improved.

I'd love to see Repeat layers to create FrankenModels by dnhkng · Pull Request #275 · turboderp/exllamav2 finally gaining traction. I think there's enough evidence by now that the self-merging actually improves performance, so by doing on the fly would let us iterate and get even better results much faster.