error loading model hyperparameters
I've been enjoying this model quite a lot - it's great. But with the latest version of llama.cpp / Wllama it won't load anymore:
https://github.com/ngxson/wllama/issues/106#issuecomment-2272304472
Just to check/hope: is there an updated version in the works?
Or are you perhaps working on a Dutch version of Gemma 2 2B it?
*crosses fingers
I have never really widely announced this model, as it was just a quick train only on instructions. So I won't spend time on updating this version.
However, I can tell that if my GPU credits allow, newer models are in the making of different sizes.
Great to hear!
For a quick quick train I was actually very impressed with how well it worked. I made it my standard Dutch model :-D
So I won't spend time on updating this version.
I learned that in theory it just needs a quick fix using a Python script to add the missing meta-data into the gguf file. It doesn't require a full recreation of the model.
In fact, with a little luck this will be fixed in Wllama, so that it can be 'hacked' in on the fly when starting inference.
Details: https://github.com/ngxson/wllama/issues/106#issuecomment-2273719761
@BoscoTheDog I tried out LM Studio yesterday and that seemed to work well, so hopefully your issues are also solved.
Yep, some default values were added to llama.cpp:
https://github.com/ngxson/wllama/issues/106#event-13928037298
I noticed you created a new version of Fietje based on Phi 2. Did you manage to get some GPU time for some newer models? I've read that Phi 3.5 is already much more multi-lingual and theoretically already supports Dutch. Although I'm pretty sure you could boost that to much greater heights.