Duplicate entry for 'YanLabs/gemma3-27b-it-abliterated-normpreserve'
There are duplicate entries for these two - with substantially different scores.
YanLabs/gemma3-27b-it-abliterated-normpreserve
YanLabs/gemma-3-27b-it-abliterated-normpreserve
The link for the first one (no hyphen) redirects to the link used for the second one. I would have guessed that the v1 variant was the same as the old one (as the author mentions having made changes to the one not labelled as v1 - but the scores for that one are substantially different as well.
These are 2 different models.
YanLabs/gemma-3-27b-it-abliterated-normpreserve
YanLabs/gemma-3-27b-it-abliterated-normpreserve-v1
The one without the hyphen (YanLabs/gemma3-27b-it-abliterated-normpreserve) is the old version of YanLabs/gemma-3-27b-it-abliterated-normpreserve. If you look at the commit history for that model, it was completely replaced on the 2nd/3rd of December 2025.
YanLabs/gemma3-27b-it-abliterated-normpreserve was tested on the 29th of November, and then the model was replaced a few days later. Then DontPlanToEnd tested the new version again under the YanLabs/gemma-3-27b-it-abliterated-normpreserve entry.
Let me clarify - there are three entries - the first one (and by extension the second one) are the ones I'm referring to. Note that the first row below does not have a hyphen in gemma3 unlike the others. (I appreciate you looking into this, much like I did, BTW! I should have been more explicit and just included a table like this the first time - whoops.)
| Model Name | URL | Release Date | Test Date |
|---|---|---|---|
YanLabs/gemma3-27b-it-abliterated-normpreserve |
https://huggingface.co/YanLabs/gemma3-27b-it-abliterated-normpreserve |
2025-11-28 | 2025-11-29 |
YanLabs/gemma-3-27b-it-abliterated-normpreserve |
https://huggingface.co/YanLabs/gemma-3-27b-it-abliterated-normpreserve |
2025-12-03 | 2025-12-31 |
YanLabs/gemma-3-27b-it-abliterated-normpreserve-v1 |
https://huggingface.co/YanLabs/gemma-3-27b-it-abliterated-normpreserve-v1 |
2025-12-03 | 2025-12-31 |
Yes, as stated above the 2nd and 3rd models in that table are 2 completely separate models, with different abliteration settings/targets.
The 1st model is the previous version of the 2nd model, that was "overwritten" by the 2nd model. The 1st model doesn't exist anymore, or rather the 2nd model is the current version of it.
Edit: I think the v1 in the model name may be confusing you. Imagine the model is called gemma-3-27b-it-abliterated-normpreserve-alternative instead of gemma-3-27b-it-abliterated-normpreserve-v1 and it might make sense.
That's roughly what I figured the case was, but I think it probably makes sense to either rename the model in this (perhaps with the date of testing?) or to remove it - but, that's a subjective decision, of course. (The model-as-tested doesn't really exist at the link, since it just redirects to the current model - though given the git/xet backend, presumably one could still access the previous version if they wanted to? Maybe the commit ID would be a better choice for the name.)
Actually, having commit IDs as unique IDs for this would probably be super helpful! But that's a whole different topic, and probably more than a small pain to implement.