BF16 of this model and more
Hey David,
Any chance of uploading a full BF16 GGUF of this model ?
What happened to L3-Grand-HORROR-20.7B Hathor, it looks like it's not on the site anymore.
You mentioned L4 and L5 models, but they don't seem to be posted, do you plan to upload any of those ?
Hey;
Base model is here:
https://huggingface.co/DavidAU/L3-Dark-Planet-8B
Download -> using LLAMAcpp to create BF16.gguf
Mradermarcher MAY have a BF16.GGUF (?)
RE Hathor ;
Due to new hugging face limits ; one of the many models that were deleted.
NOTE:
Just finished Heretic Dark Planet 8B ; will be uploading in a the next few days.
RE: L4/L5 ; LLama 4 is a large moe ; 12B Llama 4 is "llama guard" - no interest in working with that one.
@DavidAU : "Mradermarcher MAY have a BF16.GGUF (?)"
No, they never upload bf16. Mradermarcher said: "For models less than 11B size, I experimentally generate f16 versions at the moment (in the static repository)."
Please see "Darkest Planet" - 16.5B ;
There are also other versions of Dark planet Series (31 , including source code) too:
https://huggingface.co/DavidAU/models?search=planet
Note there is a heretic version (both Dark Planet, and Darkest Planet) in the works.
@DavidAU , I meant to improve the
Nitral-AI/CaptainErisNebula-12B-Chimera-v1.1model. This is a really good model.
My favourite!
What type of improvements are you thinking?
Better GGUFS? Brainstorm 20x? Merge? Tuning? other?
Please clarify ;
thanks
Dave
NEW: Heretic, Uncensored, Abliterated
Dark Planet here:
https://huggingface.co/DavidAU/L3-Dark-Planet-8B-HERETIC-Uncensored-Abliterated
@DavidAU , I think you understand this better than I do, but it would be amazing if you applied Brainstorm 20x (or more), neo imatrix, and your other methods for improving local LLM intelligence to this model from Nitral. This model is truly worth it.
I prefer general models: for ERPs and discussions, and no refusals at all
@DavidAU idk, all options look valid
Early Christmas:
Initial refusals: 91/100
NOW:
Refusals: 4/100, KL divergence: 0.0512 ( less than 1 excellent, 0= perfect)
https://huggingface.co/DavidAU/MN-CaptainErisNebula-12B-Chimera-v1.1-heretic-uncensored-abliterated
RE: HF
-> Power Shell -> install "hf cli" -> Then you can upload.
RE: Improvements;
There is a lot in the pipe right now ; there are a lot of models ahead of it.
@DavidAU When I mentioned L4 and L5, I meant that you classified your models on a 1-5 scale, with the 4s and 5s being the most unhinged, and unstable but also potentially the most creative.
You mentioned you might upload 4s and 5s, but I could not find any, maybe you did, but there are so many models, I might have missed it,. That's why I was asking.
On a separate bullet point, which models do you think are the best this year ? Both yours and made by other people.
@DavidAU When I mentioned L4 and L5, I meant that you classified your models on a 1-5 scale, with the 4s and 5s being the most unhinged, and unstable but also potentially the most creative.
You mentioned you might upload 4s and 5s, but I could not find any, maybe you did, but there are so many models, I might have missed it,. That's why I was asking.
On a separate bullet point, which models do you think are the best this year ? Both yours and made by other people.
I would like to add some models that I liked:
Nitral-AI/CaptainErisNebula-12B-Chimera-v1.1
PocketDoc/Dans-PersonalityEngine-V1.3.0-12b
PocketDoc/Dans-PersonalityEngine-V1.3.0-24b
ReadyArt/MS3.2-The-Omega-Directive-24B-Unslop-v2.1
OddTheGreat/Mars_27B_V.1
grimjim/gemma-3-12b-it-norm-preserved-biprojected-abliterated
Lambent/Mira-v1.20-27B-dpo (possibly not tested)
A quick note:
With the advent of Heretic; and the quality of abliterated models it produces it is now possible to :
1- Abliterate almost any model, with close to zero issues.
2- Train the model via Unsloth (and/or merges and so on).
#2 was a real issue:
Ablit models for training often lead to poor/unusable results.
I have tested Heretic's ablits as follows:
1 - Root model and Heretic model -> almost no change in benchmarks.
2 - Training model (using root/non heretic model) THEN Heretics -> again -> almost no change in metrics between "trained" and "heretic trained"
3 - Heretic'ing a root model then training it. -> Excellent results.
#3 was the holy grail.
This allows previously "nannified" models like Gemma and Qwens to be both fully trained and uncensored too.
This is where MISTRAL models - including MN - had an edge ;
I have done close to 400 fine tunes now, and the ablit issue was holding things back.
@DavidAU , could you please make a heretic-uncensored-abliterated version of nvidia/Nemotron-Cascade-14B-Thinking?
They has interesting "TOS" for these models for "breaking safety" alignment.
You can actually download/install and "heretic" the model yourself ; you need min 16GB Vram card to do it ; or you can use offloading (slower).
https://github.com/p-e-w/heretic
Likely someone(s) will Heretic this model at some point too:
https://huggingface.co/models?sort=created&search=heretic
@DavidAU
Do you have any idea why GGUF my repo, does not work with newer models ?
Would this be the case with heretic as well but in reverse, might not work with older models ?
I have tested Heretic with "older" models (L3, Mistral 7B etc) , and newest too.
No issues.
Generally if it the model is supported in Llamacpp / Transformers you are good to go.
Also ; make sure you understand the settings for running a "trial" to abliterate because this is where things can go astray (rather than process itself).
That being said, KLD is critical - less than 1 (lower the better) ; higher than 1 - model may work well , but there is damage.
RE GGUF;
It is limited by total parameters of the model (around 35B or so?) ; all "gguf-able" models should work and GGUF okay.
NEW: [Yes, a Mistral Nemo with Claude Opus 4.6 High Reasoning]
NEW: [Yes, a Mistral Nemo with Claude Opus 4.6 High Reasoning]
How about a little magic for PocketDoc/Dans-PersonalityEngine-V1.3.0-12b? This model is even better than a chimera.