I really like Honey Yuzu but I found Fimbulvetr v2 performs much better than v2.1 so I wanted to remake the merge with it, and I messed around with the gradient merge methods a bit too. Although I haven't tested it extensively, this version consistently scores a pointer higher than the original on a local EQ-Bench (Q4KL quant). It ties Lyra-v1, the new Nemo finetune that scored highest!
MERGE BROUGHT TO YOU BY:
- Mergekit Config Stolen From matchaaaaa/Honey-Yuzu-13B
- Trying to use lazymergekit at first except I had to throw everything out anyway and do the multi-step merge by hand
slices: # this is a quick float32 restack of BLC using the OG recipe
- sources:
- model: SanjiWatsuki/Kunoichi-7B
layer_range: [0, 24]
- sources:
- model: SanjiWatsuki/Silicon-Maid-7B
layer_range: [8, 24]
- sources:
- model: KatyTheCutie/LemonadeRP-4.5.3
layer_range: [24, 32]
merge_method: passthrough
dtype: float32
name: Big-Lemon-Cookie-11B
---
models: # this is a remake of CLC with the newer Fimbul v2.1 version
- model: Big-Lemon-Cookie-11B
parameters:
weight: 0.8
- model: Sao10K/Fimbulvetr-11B-v2 # Fim 2.1 performs significantly worse imo, and we don't care about 16k
parameters:
weight: 0.2
merge_method: linear
dtype: float32
name: Chunky-Lemon-Cookie-11B
---
slices: # 8 layers of WL for the splice
- sources:
- model: senseable/WestLake-7B-v2
layer_range: [8, 16]
merge_method: passthrough
dtype: float32
name: WL-splice
---
slices: # 8 layers of CLC for the splice
- sources:
- model: Chunky-Lemon-Cookie-11B
layer_range: [8, 16]
merge_method: passthrough
dtype: float32
name: CLC-splice
---
models: # this is the splice, a gradient merge meant to gradually and smoothly interpolate between stacks of different models
- model: WL-splice
parameters:
weight: [1, 1, 0.75, 0.625, 0.5, 0.375, 0.25, 0, 0] # 0.125 / 0.875 values removed here - "math gets screwy"
- model: CLC-splice
parameters:
weight: [0, 0, 0.25, 0.375, 0.5, 0.625, 0.75, 1, 1] # 0.125 / 0.875 values removed here - "math gets screwy"
merge_method: della_linear # New Meme
base_model: WL-splice
dtype: float32
name: splice
---
slices: # putting it all together
- sources:
- model: senseable/WestLake-7B-v2
layer_range: [0, 16]
- sources:
- model: splice
layer_range: [0, 8]
- sources:
- model: Chunky-Lemon-Cookie-11B
layer_range: [16, 48]
merge_method: passthrough
dtype: float32
name: Honey-Yuzu-Mod-13B
Meaningless EQ-Bench results at Q4KL:
This model: 78.7
matchaaaaa/Honey-Yuzu-13B: 77.64
Sao10K/MN-12B-Lyra-v1 with Mistral prompt: 78.4-78.7
senseable/WestLake-7B-v2 at Q6K: 79.15 (official EQ-Bench site reports 78.7)