kromeurus/L3-Crepuscular-Horizon-v0.1-13.3B

Well, this was hell to make. Passthrough merges aren't fun to start but this was another level of fuck you if the configs say anything. Crepus Horizon was an attempt at an expanded merge of Sunfall and Daybreak Stheno by crestf4ll, mainly for the verbosity I was able to get out of it. This and Blackened Sunfall are what I'm calling 'cracked starters', models that were purposefully broken in ways to force certain attributes at the cost of coherency, context following, etc. These models are not stand alone, at all; they're merge fodder. This one especially can misspell and struggles with anatomy, settings, and dialogue.

So, be careful. Do not use this by itself.

Quants

I don't recommend quanting this just to save your GPUs, but I can't stop you either. You've been warned though.

Details & Recommended Settings

Way more vivid and colorful language compared to both default Sunfall and Daybreak. I think I hit the censorship/Diamond Law layer somewhere somehow since it's a whole other level of descriptive. Hoping to pass that onto other merge models. Struggles to spell basic words and will misspell name half the time. At high temps, will go off on random tangents treating descriptions like dialogue and stretch out words for fun.

Rec. Settings: None, there is no making this model work.

Models Merged & Merge Theory

• crestf411/L3-8B-daybreak-stheno-v3.1
• crestf411/L3-8B-sunfall-v0.4-stheno-v3.2

What merge theory? Kidding, there is a rough direction I was going with this.

General idea was to start of with a mildly scuffed passthrough merge that kept most of the intellect and information from Daybreak and Sunfall, then try to crack it just enough to get the verbose language out. Fucking with a bunch of lists and layers, testing for about a week straight, you get the following. The first merge 'horizon1' is the more stable (barely stable) merge to come out that kept all the characteristics I was looking for while keeping coherency in mind. 'horizon3' on the other hand is the fully broken model that really pulls out the big guns.

If you're wondering where 'horizon2' is, that merge was very stable and had a bit of color to it but completely ruined the final dare_linear merge no matter what I did, so it got removed.

Configs

models:
slices:
- sources:
  - layer_range: [0, 9] # top cap
    model: crestf411/L3-8B-daybreak-stheno-v3.1
- sources:
  - layer_range: [7, 9] # top bit
    model: crestf411/L3-8B-sunfall-v0.4-stheno-v3.2
- sources:
  - layer_range: [11, 13]
    model: crestf411/L3-8B-daybreak-stheno-v3.1
- sources:
  - layer_range: [8, 13]
    model: crestf411/L3-8B-sunfall-v0.4-stheno-v3.2
- sources:
  - layer_range: [15, 16]
    model: crestf411/L3-8B-daybreak-stheno-v3.1
- sources:
  - layer_range: [10, 16] # mids
    model: crestf411/L3-8B-sunfall-v0.4-stheno-v3.2
- sources:
  - layer_range: [16, 24]
    model: crestf411/L3-8B-daybreak-stheno-v3.1
    parameters:
      scale:
      - filter: k_proj
        value: 0.8
      - filter: down_proj
        value: 0.8
- sources:
  - layer_range: [16, 24]
    model: crestf411/L3-8B-sunfall-v0.4-stheno-v3.2
- sources:
  - layer_range: [24, 26] # end bit
    model: crestf411/L3-8B-daybreak-stheno-v3.1
    parameters:
      scale:
      - filter: v_proj
        value: 0.8
- sources:
  - layer_range: [27, 29]
    model: crestf411/L3-8B-sunfall-v0.4-stheno-v3.2
    parameters:
      scale:
      - filter: v_proj
        value: 0.8
- sources:
  - layer_range: [29, 31]
    model: crestf411/L3-8B-daybreak-stheno-v3.1
- sources:
  - layer_range: [23, 32] # end cap
    model: crestf411/L3-8B-sunfall-v0.4-stheno-v3.2
parameters:
  int8_mask: true
merge_method: passthrough
name: horizon1
---
models:
slices:
- sources:
  - layer_range: [0, 12]
    model: crestf411/L3-8B-daybreak-stheno-v3.1
- sources:
  - layer_range: [3, 8]
    model: crestf411/L3-8B-sunfall-v0.4-stheno-v3.2
- sources:
  - layer_range: [5, 10]
    model: crestf411/L3-8B-daybreak-stheno-v3.1
- sources:
  - layer_range: [9, 13] 
    model: crestf411/L3-8B-sunfall-v0.4-stheno-v3.2
- sources:
  - layer_range: [12, 15]
    model: crestf411/L3-8B-daybreak-stheno-v3.1
- sources:
  - layer_range: [16, 19]
    model: crestf411/L3-8B-sunfall-v0.4-stheno-v3.2
- sources:
  - layer_range: [20, 26]
    model: crestf411/L3-8B-daybreak-stheno-v3.1
- sources:
  - layer_range: [22, 29]
    model: crestf411/L3-8B-sunfall-v0.4-stheno-v3.2
- sources:
  - layer_range: [26, 31]
    model: crestf411/L3-8B-daybreak-stheno-v3.1
- sources:
  - layer_range: [26, 32]
    model: crestf411/L3-8B-sunfall-v0.4-stheno-v3.2
parameters:
  int8_mask: true
merge_method: passthrough
name: horizon3
---
models: 
  - model: parts/horizon1
    parameters:
      weight: [0.9, 1, 0.75]
  - model: parts/horizon3
    parameters:
      weight: [0.1 , 0, 0.25]
merge_method: dare_linear
parameters:
  normalize: false
  int8_mask: true
base_model: parts/horizon1
dtype: bfloat16