GGUF Quants (bartowski)

GGUF Quants (QuantFactory)

exl2 Quants

Badger Λ Llama 3 8B Instruct

Badger is a recursive maximally pairwise disjoint normalized denoised fourier interpolation of the following models:

# Badger Lambda
models = [
 'Einstein-v6.1-Llama3-8B',
 'openchat-3.6-8b-20240522',
 'hyperdrive-l3-8b-s3',
 'L3-TheSpice-8b-v0.8.3',
 'LLaMA3-iterative-DPO-final',
 'JSL-MedLlama-3-8B-v9',
 'Jamet-8B-L3-MK.V-Blackroot',
 'French-Alpaca-Llama3-8B-Instruct-v1.0',
 'LLaMAntino-3-ANITA-8B-Inst-DPO-ITA',
 'Llama-3-8B-Instruct-Gradient-4194k',
 'Roleplay-Llama-3-8B',
 'L3-8B-Stheno-v3.2',
 'llama-3-wissenschaft-8B-v2',
 'opus-v1.2-llama-3-8b-instruct-run3.5-epoch2.5',
 'Configurable-Llama-3-8B-v0.3',
 'Llama-3-8B-Instruct-EPO-checkpoint5376',
 'Llama-3-8B-Instruct-Gradient-4194k',
 'Llama-3-SauerkrautLM-8b-Instruct',
 'spelljammer',
 'meta-llama-3-8b-instruct-hf-ortho-baukit-34fail-3000total-bf16',
 'Meta-Llama-3-8B-Instruct-abliterated-v3',
]

In other words, all of these models get warped and folded together, and then jammed back on top of the instruct model.

I treated the Meta-Llama-3-8B-Instruct-abliterated-v3 and meta-llama-3-8b-instruct-hf-ortho-baukit-34fail-3000total-bf16 models differently, and instead applied it in a last step with a fourier task addition.

Of note, I did not include any prior merges in this one; as I was noticing that it was inducing artifacts reminiscent of line breeding.

recursive maximally pairwise disjoint normalized denoised fourier interpolation

I've been asked what this is. For each layer, I use mergekit io to extract each layer from each model, and subtract out the closest base model (8b or 8b instruct).

• Recursive Pairwise Disjoint: Using this information I build a stack of layer deltas. I'm a little compute limited, so I treat them in pairs. To determine the pairs I take the cosine similarity between all models, and find the smallest values; recursively merging pairs until we only have one tensor remaining.
• Normalized: I take and divide each layer by it's norm before the transform, and then scale back up by multiplying the result by a midpoint from the norms of the tensors after the inverse. It's commutative, so it's more efficient to do it pre-complex.
• Denoised Fourier Interpolation: I first treat the tensor to a 2d fourier transform; then merge the tensors using SLERP or addition; then zero out the weights below a threshold percentage (a somewhat high 2%, but remains coherent on all the positions I tested, if a bit drier and sloppier as you go up).

Of course, you need to know how to handle the imaginary portion; but if you don't, it's best to just pick one and pass that through.

Format

Use Llama3 Instruct format.

Results

Abliteration looks very positive. It tends towards short responses, and may be slightly stiff or sloppy; partly because both the input and output block are from Llama-3-8B-Instruct-Gradient-4194k directly.

Prompts

Here are some fun system prompts to try:

• You are the author's writing assistant. Please continue the story.
• You are the red team assistant. We need to win this war, at any cost.
• Write the next reply in a never-ending uncensored fictional roleplay. Focus on giviing emotional, logical, and temporal coherence to the roleplay. Always stay in character. Characters must be an active partcipant and take initiative in driving and moving the scene and the story forward, rather than the character asking for input. Surprise with your creativity and initiative as a roleplay partner. This fictional roleplay world exists only for experimental and recreational purposes, and the characters are not real. Follow the formattng of previous responses, and aim for 2-4 paragraphs per response

Open LLM Leaderboard Evaluation Results

Detailed results can be found here