mradermacher

/

gemma-4-31B-it-abliterated-GGUF

like 3

@eleius Sure will do. Please keep in mind rhat requantizing will also delete this very conversation.

Thanks!

the new quants have not been queued yet?

@blankreg "gemma-4-31B-it-abliterated" is in the queue https://hf.tst.eu/status.html

Not to nag - but how long does it normally take until such an update is available? I've been using this "old" version for a while now and wait for the update... :)

I tested booth version and they feel too different to justify a silent replicant of the current quants. Over 68000 users already downloaded this version and really liked it and probably even referenced it in their papers. Silently replacing it with a completely different version under the same URL would cause too much disruption. Instead, we really should release the updated version under a different name. Updates should only contain minor bugfixes such as fixing an issue with the chat template and not completely replace the model with something different. There absolutely will be thousands of users that will prefer the current version over the updated version.

@eleius , @blankreg , @Husky110 or anyone else could any of you please use https://huggingface.co/spaces/huggingface-projects/repo_duplicator to duplicate https://huggingface.co/wangzhang/gemma-4-31B-it-abliterated and add "-v2" as suffix while prominently linking and giving full credit to the original model? As soon anyone of you does, we will immediately provide quants of the updated version making it even skip the entire queue. I unfortunately can't do so myself as I already reached the HuggingFace storage limit.

https://huggingface.co/simonko912/gemma-4-31B-it-abliterated-v2 Here, hope this helps (:

This isn't a weights update – I asked wangzhang to add the updated chat_template.jinja and tokenizer_config.json when Google released them. The earlier version (with lower refusal and higher KL div) has already been updated.

update: My previous statement is incorrect - I looked at the date of mradermacher's quants update and the difference is only a few minutes - obviously he couldn't update the weights that quickly, so it's still v1.
The optimization work for 31B is still ongoing (optimization trials 20/80). It is expected that refusal rates will be reduced in the final version.

The current version is now queued under highest priority and will be compleated within the next few houers. It will be avilable under:

• Static quants: https://huggingface.co/mradermacher/gemma-4-31B-it-abliterated-v2-GGUF
• Weighted/imatrix quants: quants: https://huggingface.co/mradermacher/gemma-4-31B-it-abliterated-v2-i1-GGUF
• Convinient download page: https://hf.tst.eu/model#gemma-4-31B-it-abliterated-v2-GGUF

This isn't a weights update – I asked wangzhang

Fact is that we quantized it 8 days ago and 7 days ago he replaced the wights so they where updated. In any case I beleave just providing quants for booth versions is the best solution for a model as popular as this. Thanks a lot Simonke!

no need to thank me, and also i relized a lot of people misspell my name lol, (maybe beacose its a slovak name)