More experiments that actually work LMAO. Started straying away from Siithamo, at least model list wise. Stheno is just so chatty, idk how to tame it yet. Used components from my upcoming fatboy model as parts of this merge and imo, this is a hidden gem.
09.27.2024: Well this model needed way more help then i thought it did. So a month later, I come with a new iteration Aglow Vulca. Use that one, since this has bugs.
Quants
OG Q8 GGUF by me.
Model Details & Recommended Settings
(Still testing; details subject to change)
Follows instructs fairly well, doesn't stray much unless the temp is too high. Same thing as all the other model I make with Formax (ty ArliAI), this merge will reflect the character card quality; shit card will have shit output and vise versa.
Generates slightly flowery text, thought process type writing. Human-ish dialogue. Chatty but not too chatty, will mimic previous text examples. Coherent up to 16k (as tested).
Rec. Settings:
Template: L3
Temperature: 1.3
Min P: 0.1
Repeat Penalty: 1.05
Repeat Penalty Tokens: 256-512 #stick closer to 256
Merge Theory
Will update later, too tired rn.
Config
models:
- model: ArliAI/ArliAI-Llama-3-8B-Formax-v1.0
parameters:
weight: [1, 1, 1, 1, 0, 0, 0, 0]
- model: gradientai/Llama-3-8B-Instruct-Gradient-1048k
parameters:
weight: [0, 0, 0, 0, 1, 1, 1, 1]
base_model: ArliAI/Llama-3.1-8B-ArliAI-Formax-v1.0
parameters:
normalize: false
int8_mask: true
merge_method: dare_linear
dtype: float32
out_dtype: bfloat16
tokenizer_source: base
name: formaxext.3.1
---
models:
- model: kromcomp/L3-Ceto-Epith-Humanity.A-v0.1-8B
- model: ghost-x/ghost-8b-beta-1608
- model: kromcomp/L3-Ceto-Epith-Humanity-v0.1-8B
base_model: tohur/natsumura-storytelling-rp-1.0-llama-3.1-8b
parameters:
normalize: false
int8_mask: true
merge_method: model_stock
dtype: float32
out_dtype: bfloat16
tokenizer_source: base
name: humplus
---
models:
- model: humplus
parameters:
weight: [0.01, 0.53, 0.9]
- model: ArliAI/Llama-3.1-8B-ArliAI-RPMax-v1.1
parameters:
weight: [0.55, 0.29, 0.1]
- model: crestf411/L3.1-8B-sunfall-v0.6.1-dpo
parameters:
weight: [0.54, 0.28, 0.1]
base_model: humplus
parameters:
normalize: false
int8_mask: true
merge_method: dare_linear
dtype: float32
out_dtype: bfloat16
tokenizer_source: base
name: tusl3.1
---
models:
- model: tusl3.1
parameters:
weight: [0.5, 0.75, 0.8, 0.9, 0.95]
density: 0.9
gamma: 0.01
- model: formaxext.3.1
parameters:
weight: [0.5, 0.25, 0.2, 0.1, 0.05]
density: 0.9
gamma: 0.01
base_model: tusl3.1
tokenizer_source: union
parameters:
normalize: false
int8_mask: true
merge_method: breadcrumbs_ties
dtype: float32
out_dtype: bfloat16
name: mantusl3.1