mradermacher/model_requests · Update for Zurich Models & Addition of Geneva Models

3 days ago

Hi again!

I was reviewing Zurich 7B and 14B and I noticed I made a mistake in its tokenizer configuration and that it wasn't optimised to my training dataset chat template format, and this causes them to hallucinate more and underperform. I corrected the mistake, but I noticed it was also present in the quantized GGUF files after I reviewed their file metadata. The mistake only affects the 7B and 14B models, the 1.5B model is completely unaffected.

I apologize for the massive hassle.

These are the following model repos with the incorrect tokenizer (all 14B and 7B variants with different sizes):

I also would like to request the Geneva 12B models be quantized, these are the follow versions:

https://huggingface.co/rubenroy/Geneva-12B-GCv2-10k
https://huggingface.co/rubenroy/Geneva-12B-GCv2-50k
https://huggingface.co/rubenroy/Geneva-12B-GCv2-100k
https://huggingface.co/rubenroy/Geneva-12B-GCv2-500k
https://huggingface.co/rubenroy/Geneva-12B-GCv2-1m
https://huggingface.co/rubenroy/Geneva-12B-GCv2-5m

mradermacher

Owner 2 days ago

First, thanks for listing all the affected repos, that was helpful.

I've queued the geneva models (in fact a while earlier, they should already be done), and deleted and requeued the 7b/14b Zurich models, however, given the stawte of the queue at the moment, it can take a bit before they are all through.

I am relieved that it's "just" <=14b's. If they were much larger, I'd probably strangle you by now. Not because you specifically caused this issue, but because we are already tight for a multitude of reasons at the moment, and you'd be a convenient victim coming along at the right time :)

Anyway, cheers!

mradermacher changed discussion status to closed 2 days ago