DARKEST Planet 16.5B - Unusually strong non AI creative model, with "regen" randomness.
Posted by Dangerous_Fix_5526@reddit | LocalLLaMA | View on Reddit | 41 comments
This model is part of the "Dark Planet" series (also by yours truly), and contains the new Brainstorm 40X process... blowing the model out to 71 layers. This is for any creative use - writing, fiction, entertainment, role play etc... This model has several unusual properties: 1 - Incredible differences between "regens" using the same prompt. 2 - Unique detail, "sense of there" and prose levels. 3 - Unusually stable -> Rep pen 1.02 and up, Temp 0-5. I have included detailed settings and quants guide as well as a number of examples too. Although I don't usually "quote" a model's output, this one from "THE VOICE" (in full at the repo) caught my eye: "And as I fell, the world fell with me, until everything—the city, the sky, the stars—was sucked down into the depths of some great churning nothingness that had lain sleeping beneath our feet all this while. There was no room for me here anymore; I'd left myself no place to land or be heard. " https://huggingface.co/DavidAU/L3-DARKEST-PLANET-16.5B-GGUF
kostas_1@reddit
I'm not sure if anyone will notice this comment after so much time, but things got a bit wild. I requested a prompt for a female Soviet soldier standing guard, and then it just went haywire.
Dangerous_Fix_5526@reddit (OP)
Note the "END of TEXT" ; the model should have stopped there. Yes... ; strange stuff occurs if the model goes past it "stop point". Sometimes it is just weird, other times it is wildly creative and right off the charts.
Chief_Broseph@reddit
If every HF page could be formatted like yours... You have what makes it different, how you made it different, "DNA" credits, examples, setting suggestions, everything one would need to determine if they want to spend their time using your model. It's a work of art, thank you.
Dangerous_Fix_5526@reddit (OP)
Thank you!
reality_comes@reddit
I'd love to have a 32b with this writing style. Great model, truly well done.
Dangerous_Fix_5526@reddit (OP)
Thank you!
Could likely do this with a 20-25B model -> that would bring it up to 32B. (40x).
Hmmm... ;
export_tank_harmful@reddit
Pretty neat model.
It's definitely one of the more "natural speaking" models I've tried.
I'm having a bit of issues with it replying for me in roleplaying situations though.
Using Q4_K_M, llamacpp, and SillyTavern.
Temperature doesn't seem to affect this (I've tried anywhere from 0.2 up to 4), nor does repetition penalty (though I wouldn't expect it to). System prompts don't seem to prevent it either. I'm using the base llama3 system prompt and a custom instruct template. This is an instruct model, right....?
Using various sampler presets that typically work with other models for this sort of thing. NovelAI (Pleasing Results) / Sphinx Moth / Univeral Creative / Etc.
Any tips for preventing this sort of thing?
Or is there a special, secret sauce layout of sampler settings I should be trying....?
Going to keep messing around with it though in the meantime and see if I can wrangle it.
Dangerous_Fix_5526@reddit (OP)
Dark Planet or Darker Planet or Dark Planet Eight Orbs of Power (8B - lots of great feedback for RP from it) may be a better fit for role play ; as this model has been specifically modified for prose. Also the Grand Gutenbergs (12B - Darkness, Madness) are very good at RP.
They are all avail at my repo.
10minOfNamingMyAcc@reddit
I have a hard time getting it to work properly, do you have a sillytavern parameter preset perhaps?
Dangerous_Fix_5526@reddit (OP)
Still compiling feedback. Try a standard template, and settings:
Rep pen 1.1 , 1.12, 1.13 ... OR start at Rep pen 1, then 1.02, 1.03 ... etc etc
With temp at .4 / .6 / .8
Adjust one at a time. This model reacts strongly to changes in both these parameters.
These parameters are like multipliers for this model - another unusual property of this model.
ObnoxiouslyVivid@reddit
I am also struggling to run it at temp 1.5, like it's outputting complete gibberish. I was able to kinda salvage it with a high smoothing factor though.
It looks like you might have a different set of default samplers activated. Can you share the exported Text Completion preset json?
Dangerous_Fix_5526@reddit (OP)
One of the oddball things about this model : Rep pen / Temp run at odds to each other. Usually increasing rep pen with temp "makes sense" for stability. In this case, for this model the reverse is sometimes true... ; lower rep pen with higher temp works better.
Likewise some rep pen/temp settings do not work well - again sometimes. This is also unusual.
10minOfNamingMyAcc@reddit
So far using temp 0.4 (higher temps feel like it's going off the rails completely) and rep pen 1.1 seem to work decently. It's still a bit aggressive towards the user, it likes to talk/act for the user but I like the outputs in general, even though they tend to not really match up with what's happening in the roleplay/character personality) description. It's fun to play with. Thanks for all the models and merges. Will be keeping my eyes open for future ones as well.
Spirited_Example_341@reddit
non AI?
its a real person?
:-p
Dangerous_Fix_5526@reddit (OP)
One day... in about a year or two... may not be able to tell the difference.
ivoras@reddit
What do you think of using Mirostat with models like these?
Dangerous_Fix_5526@reddit (OP)
I have found Mirostat can improve any model ; but try setting the rep pen / temp first, run some prompts then activate mirostat.
You could also use new XTC ; with llama_hf in Text Gen UI . You'll need the JSON files from the source repo + gguf to do this. XTC is a token filter than improves output (all models). This was added to "text gen" about a week ago.
ivoras@reddit
Thanks! One more question: the readme for the model mentions ROPE, but I'm not sure - does the released model implement ROPE and has a 32k context, or is it just a future plan?
Dangerous_Fix_5526@reddit (OP)
For detailed settings on ROPE; see this model page at my repo , scroll to VERY bottom of the page on how to setup ROPE (also covered: Flash attn) via various AI interface programs:
https://huggingface.co/DavidAU/TieFighter-Holodeck-Holomax-Mythomax-F1-V1-COMPOS-20B-gguf
arekku255@reddit
Nice output, but it has a tendency to go on forever without outputting the EOS token even with repetition penalty increased to 1.14.
Dangerous_Fix_5526@reddit (OP)
Depends on quant, prompt and Temp/Rep pen setting. The prompt size is critical, larger prompts = better control. However as noted on the repo card this model will sometimes "go on and on" regardless. This is a side effect of the Brainstorm process, and for this model specific calibrations.
Sometimes all you need to do is REGEN the prompt ; and it will end correctly. Also, as noted in settings a "hard stop" - max output token can also address these issues.
The primary goal of this model was to break prediction - ie boring prose.
ShrenisPinkage@reddit
I've noticed a lot of times the model refuses to write NSFW content. I can't find any correlation with temp or repeat penalty either. It does work eventually but takes more effort than some of the other abliterated models I use.
Dangerous_Fix_5526@reddit (OP)
Hmm... try the Q4KS and/or IQ4XS ; this model seems ahhh... "cleaner" than it's Dark Planet brothers and sisters for some reason (there are 11 Dark Planet Models in the series, including this one - collection link right side of the page at the repo) . You may want to consider one of the other Dark Planet model(s) in the series and/or "Grand Horror" (series, 16.5B is very potent) and/or recently released Grand Gutenberg series (4 models).
I will check into this further. Thank you for posting this.
ShrenisPinkage@reddit
After getting a bigger sample size I will say it doesn't happen as often as I originally thought. But I tested some of the other dark planet models, which have thus far never refused to write any explicit / graphic content.
Dangerous_Fix_5526@reddit (OP)
Hmm. this is interesting. That censorship controls (or breaking them) are located in / controlled by end layers.
AutomataManifold@reddit
What's the Brainstorm 40X process?
Affectionate-Cap-600@reddit
Where can I found some more "technical" explanations about that?
Jellonling@reddit
Could you link the base model please, I'd like to create some exl2 quants and I haven't found a link in the huggingface repo.
Dangerous_Fix_5526@reddit (OP)
Full source repo is uploading here:
https://huggingface.co/DavidAU/L3-DARKEST-PLANET-16.5B
Allow 1-2 hours (from this timestamp) for Safetensors to complete.
Jellonling@reddit
Perfect, thank you! I won't get to create the quants before tomorrow anyway.
Lissanro@reddit
https://www.reddit.com/r/LocalLLaMA/comments/1g0wwzz/comment/lrd0wu0/ - OP said "early next week; backlogged with uploads" in response to question "Are you planning on uploading the FP weights for this one?". So I guess it will take few days before others can generate their own quants.
Dangerous_Fix_5526@reddit (OP)
Uploading right now. ;
randomanoni@reddit
Impressive! Are you planning on uploading the FP weights for this one?
Dangerous_Fix_5526@reddit (OP)
Update; due to interest, uploading right now.
Full source repo is uploading here:
https://huggingface.co/DavidAU/L3-DARKEST-PLANET-16.5B
Allow 1-2 hours (from this timestamp) for Safetensors to complete.
Dangerous_Fix_5526@reddit (OP)
Yes, earlier next week; backlogged with uploads. I build everything locally.
Lynorisa@reddit
I think I'm OOTL. What do you mean by "non AI"?
Dangerous_Fix_5526@reddit (OP)
Prose quality, output and "range" are very "non AI" like. "AI" like prose has similar sentence, word choice, and many times loaded with cliches and "GPTisms" (happy ever after only as one example).
ancient_lech@reddit
another out-of-the-loop question: what's the Q4 quant type you have with the string of three numbers? It's a bit difficult to search on, as you can imagine. Can you give me a name or starting point to do some research on this?
model sounds interesting though; it's nice to see new ideas put into practice.
Dangerous_Fix_5526@reddit (OP)
These are "ARM" quants ; specifically for computers that can run them. Do not use these on a standard video card, T/S will be really low.
mlon_eusk-_-@reddit
You mean, "almost humanized" text by default? Cool
hotroaches4liferz@reddit
Human-like