MarsupialAI

Q4KS GGUF for https://huggingface.co/MarsupialAI/Dumbstral-169B No imat, no other quant schemes. This is all I'm willing to do for a model that nobody can reasonably run. FSM help Bartowski and Mradermacher if they choose to run full quant sets for this bastard.

NaNK

—

KitchenSink_103b

NaNK

llama

Yeet 51b 200k

This model is a rotating-stack merge of three Yi 34b 200k models in a 51b (90 layer) configuration. My reasoning behind this merge was twofold: I'd never seen a stacked merge made from 34b models, and I thought that maybe this could give near-70b performance, but with a much larger context window while still fitting within 48GB of VRAM. I think the results are quite good. The model performs on par with many 70b models at RP, chat, and storywriting. At Q4KS it will fit into a pair of 24GB GPUs with 32k context. Coherency at 32k is excellent, and will probably remain very good well beyond that thanks to the 200k base training. The gotcha here is speed. While it inferences as you'd expect for the model size, it's much slower than a similarly-sized 8x7b MoE. And while I personally find the output of this model to outperform any mixtral finetune I've seen so far, those finetunes are getting better all the time, and this really is achingly slow with a lot of context. I'm getting less than half a token per second on a pair of P40s with a full 32k prompt. But that's not to say this model (or even the 51b stack concept) is useless. If you're patient, you can get extremely good output with very deep context on attainable hardware. There are undoubtedly niche scenarios where this model or similarly-constructed models might be ideal. Component models for the rotating stack are - adamo1139/Yi-34B-200K-AEZAKMI-v2 - brucethemoose/Yi-34B-200K-DARE-megamerge-v8 - taozi555/RpBird-Yi-34B-200k This model is uncensored and capable of generating objectionable material. However, it is not an explicitely-NSFW model, and it has never "gone rogue" and tried to insert NSFW content into SFW prompts in my experience. As with any LLM, no factual claims made by the model should be taken at face value. You know that boilerplate safety disclaimer that most professional models have? Assume this has it too. This model is for entertainment purposes only. FP16 and Q4KS GGUFs are located here: https://huggingface.co/MarsupialAI/Yeet51b200kGGUFQ4KSFP16 Prompt format Seems to work fine with Alpaca prompts. Considering the variety of components, other formats are likely to work to some extent. WTF is a rotating-stack merge? Inspired by Undi's experiments with stacked merges, Jeb Carter found that output quality and model initiative could be significantly improved by reversing the model order in the stack, and then doing a linear merge between the original and reversed stacks. That is what I did here. I created three passthrough stacked merges using the three source models (rotating the model order in each stack), and then doing a linear merge of all three stacks. The exact merge configs can be found in the recipe.txt file.

NaNK

llama