Please wait...
derp, I just reread the docs. the instruct data is not trade secret in this case.
related concept I ran across while browsing huggingface models: model merging.
this tries to explain it: https://developer.nvidia.com/blog/an-introduction-to-model-merging-for-llms/
I've seen some merged models in the 15B size range (runnable on gaming GPUs) that apparently outperform their base models, but you kinda have to worry about benchmaxxing...
In some cases the training data is also freely shared. True, not all, but some.
While i could be wrong, my understanding is the 'removing the rails' searches the model and looks for refusal tags and directly changes the model weights, its not like a retrain event. But regardless of how its done, in raw tests i have not seen them improve, and often not as good, but of course in tests where refusals got in the way, i can see it being better. In my own experience, i have had a 'core' model that would do x for me. I tried the abliteratied version, and it was hallucinating and rambling to the point it was unusable.
My use case is mostly to get around stupid stuff without having to con it. Example was wanting to do some API work "i cant answer that, as that could cause harm to the xyz system" finally got around it with some prompt changes, but it was rather annoying to have to do so for something that was so mundane. Another time, when i was having it do lyrics "oh, i cant create depressing lyrics .. bla bla " it was freaking goth ...that is the point.
I was probably conflating. The improved models I definitely remember seeing are merges of multiple base models.
me. I tried the abliteratied version, and it was hallucinating
and rambling to the point it was unusable.
Good to know. I have no use case for uncensored stuff so it makes sense for me at least to stay away from all that. More gernally, Huggingface leaderboards have a filter you can apply that says "original model vendor only" or words to that effect, which definitely excludes questionable things like merged models, but also excludes useful things like format conversions to OnnxRuntime.
(* format conversions can also be questionable because of the available operators and other issues, but this is clearly not the samt thing)
Yes, merges can be better.
Sat Nov 22 2025 00:39:03 UTC from LoanShark
I was probably conflating. The improved models I definitely remember seeing are merges of multiple base models.
there's so much about this field I don't understand. Like, how is it possible to merge a model from the Qwen series which something else from a different vendor that might have different I/O format conventions (e.g. left padding vs right padding just to pick one example) ...