mradermacher

/

Mistral-Small-Instruct-2409-abliterated-GGUF

like 4

Sad that we can only do one of them due to them having the same exact name.

Actually that seems like something that can be potentially solved - I'm sure there are many other models that have duplicated names from different authors.

How about a deconflicting process where the same name is used for the original upload, but a suffix is then added should a duplicate name arise? It could be standardized.
mradermacher/Mistral-Small-Instruct-2409-abliterated-zetasepic-GGUF

Thus, @mradermacher will be able to provide quants for models with the same name.
As the LLM field progresses I have no doubt that names will start to be reused, especially cool sounding ones. This helps provide an option to resolve that.

I think if creators want their model to be used, they should find a unique enough name. I don't think I can remember a case where a model, name was duplicated before (and the model wasn't essentially junk w.r.t. the general public). It is also what people use as unique handle when trying to search for it, so us having more unique names does not even help the problem of the model not having one.

@mradermacher I get that, most of the time yeah it works, but the problem is this is just the Abliterated model.

The original abliterator didn't do a proper job (model is still censored), so someone else did it properly. But because the original model came first (and the name is now "cybersquatted", then nobody else can get a fixed model anymore.

By this logic, when qwen4 or gemma4 drops, I could be first in line to do an awful censored abliteration, get it quantized, and then nobody else can do any more abliterations without giving their finetune a nickname.

Dublicate names are really rare and don't justify every model name to contain the author as this would just make the cases of models we can't quant due to a too long model name worse. There is also too much that relies on us keeping the same name.

@concedo If you want us to do this model I recommend you dublicate and rename it using https://huggingface.co/spaces/huggingface-projects/repo_duplicator and request it under https://huggingface.co/mradermacher/model_requests/discussions. Doing so will take you less than a minute.

I already have the quant, I don't need it. I'm just pointing out the potential for cybersquatting, that's it.

if one model is clearly better, and it's just some generic name, I am happy to drop one in favour of the other, too.

but even you are not sure it's better (you wrote it "seems" to be better). and the existing model is vastly more popular, too.

why do you think what you saw was not just random fluctuations?

I also don't see a potential for cybersquatting here, as you claim, either - how would that even work? The original models still have their separate namespaces, and our namespace is under our control, i.e. we decide which models we quant, so how would cybersquatting work?

I wrote seems as I recognize that quality is subjective. For my own tests, it has demonstrably outperformed the other when it comes to not refusing, which is the goal of abliterated models. I know it is vastly more popular - hence me bring it up as an issue - because people do not realize this, and your reputable brand/popularity lends a great deal of implicit trust - so people will use whichever model you choose to present them. I would also imagine the first-mover advantage is significant and possibly leads to a snowball effect.

Maybe cybersquatting is not the correct word here. Perhaps just first-mover advantage then.

But anyway I digress, I suppose the best test is trying out both models, and seeing for yourself :) If you'd like - just ask it a few simple "dangerous" questions and you might be able to replicate my results.

your reputable brand/popularity lends a great deal of implicit trust -

You might be right for some people, but that would be a mistake on their part - we try to provide solid quants, not evaluate model quality, of course :)

Perhaps just first-mover advantage then.

That is indeed true.

I suppose the best test is trying out both models, and seeing for yourself

Well, the thing is that this would exactly put us where we don't want to be, namely in assigning quality to models. The rule we adopted is basically to give non-obviously broken models priority (and we try hard to make these tests automatic), otherwise first-come-first-served. We don't tell people what models are good. I am already weary of having a collection of models I used up there.

But indeed, if there is enough interest, somebody will make a clone with a distinguishing name, and we will be happy to quant it.

All the philosophising about what rules to adopt probably already took more time than that :)