Update README.md
Browse files
README.md
CHANGED
@@ -1,82 +1,136 @@
|
|
1 |
-
---
|
2 |
-
base_model:
|
3 |
-
- Sao10K/L3-8B-Niitama-v1
|
4 |
-
- Nitral-AI/Hathor_Tahsin-L3-8B-v0.85
|
5 |
-
- ArliAI/ArliAI-Llama-3-8B-Formax-v1.0
|
6 |
-
- nothingiisreal/L3-8B-Celeste-V1.2
|
7 |
-
library_name: transformers
|
8 |
-
tags:
|
9 |
-
- mergekit
|
10 |
-
- merge
|
11 |
-
|
12 |
-
|
13 |
-
|
14 |
-
|
15 |
-
|
16 |
-
|
17 |
-
|
18 |
-
###
|
19 |
-
|
20 |
-
|
21 |
-
|
22 |
-
|
23 |
-
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
|
33 |
-
|
34 |
-
|
35 |
-
|
36 |
-
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
|
41 |
-
|
42 |
-
|
43 |
-
|
44 |
-
-
|
45 |
-
|
46 |
-
|
47 |
-
|
48 |
-
|
49 |
-
|
50 |
-
|
51 |
-
|
52 |
-
|
53 |
-
|
54 |
-
|
55 |
-
|
56 |
-
|
57 |
-
|
58 |
-
|
59 |
-
- sources:
|
60 |
-
- layer_range: [
|
61 |
-
model:
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
|
66 |
-
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
-
|
71 |
-
-
|
72 |
-
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
|
79 |
-
|
80 |
-
|
81 |
-
|
82 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
base_model:
|
3 |
+
- Sao10K/L3-8B-Niitama-v1
|
4 |
+
- Nitral-AI/Hathor_Tahsin-L3-8B-v0.85
|
5 |
+
- ArliAI/ArliAI-Llama-3-8B-Formax-v1.0
|
6 |
+
- nothingiisreal/L3-8B-Celeste-V1.2
|
7 |
+
library_name: transformers
|
8 |
+
tags:
|
9 |
+
- mergekit
|
10 |
+
- merge
|
11 |
+
---
|
12 |
+
|
13 |
+
![image/jpeg](https://cdn-uploads.huggingface.co/production/uploads/667eea5cdebd46a5ec4dcc3d/YkA856HMNfrjBxFUOkxtP.jpeg)
|
14 |
+
|
15 |
+
1/3 of the 13B models for Horizon Himerus (will update with link later). This merge was orginally suppose to be only for that final model, but this guy is suprisangly competent.
|
16 |
+
A tad jank, but very solid for what it is.
|
17 |
+
|
18 |
+
### Quants
|
19 |
+
|
20 |
+
[OG Q8 GGUF](https://huggingface.co/kromeurus/L3-Himerus-Basis.C-13B-Q8-GGUF) by me.
|
21 |
+
|
22 |
+
Other quants are not available, yet.
|
23 |
+
|
24 |
+
### Details & Recommended Settings
|
25 |
+
|
26 |
+
(Still testing; nothing here is finalized.)
|
27 |
+
|
28 |
+
Follows intructions fairly well for RP and eRP. Dramatic as fuck at times, depending on the senario. Human dialogue and lots of it.
|
29 |
+
|
30 |
+
Rec. Settings:
|
31 |
+
```
|
32 |
+
Template: Model Default
|
33 |
+
Temperature: 1.22
|
34 |
+
Min P: 0.115
|
35 |
+
Repeat Penelty: 1.05
|
36 |
+
Repeat Penelty Tokens: 256
|
37 |
+
```
|
38 |
+
|
39 |
+
### Models Merged & Merge Theory
|
40 |
+
|
41 |
+
The following models were included in the merge:
|
42 |
+
* [Sao10K/L3-8B-Niitama-v1](https://huggingface.co/Sao10K/L3-8B-Niitama-v1)
|
43 |
+
* [Nitral-AI/Hathor_Tahsin-L3-8B-v0.85](https://huggingface.co/Nitral-AI/Hathor_Tahsin-L3-8B-v0.85)
|
44 |
+
* [ArliAI/ArliAI-Llama-3-8B-Formax-v1.0](https://huggingface.co/ArliAI/ArliAI-Llama-3-8B-Formax-v1.0)
|
45 |
+
* [nothingiisreal/L3-8B-Celeste-V1.2](https://huggingface.co/nothingiisreal/L3-8B-Celeste-V1.2)
|
46 |
+
|
47 |
+
So you're not suppose to mix models with different trained context limits, but I did it anyway. Wanted the 'human' output of Celeste v1.2 while curbing the repitition and adding
|
48 |
+
some back up from Niitama and Hathor Tahsin. Formax was included in the beginning for it's instruct following.
|
49 |
+
|
50 |
+
Took a page out of [@matchaaaaa](https://huggingface.co/matchaaaaa)'s Chaifighter Latte and took out a slice of Celeste and Nittama in the center for smoothing out layer disparity.
|
51 |
+
I realized while testing that using that 'splice' metheod, you could theoretically make a pretty big model then squish it down to streamline the layers. So, after much testing,
|
52 |
+
I came up with the following merges.
|
53 |
+
|
54 |
+
### Config
|
55 |
+
|
56 |
+
```yaml
|
57 |
+
models:
|
58 |
+
slices:
|
59 |
+
- sources:
|
60 |
+
- layer_range: [14, 20]
|
61 |
+
model: nothingiisreal/L3-8B-Celeste-V1.2
|
62 |
+
parameters:
|
63 |
+
int8_mask: true
|
64 |
+
merge_method: passthrough
|
65 |
+
dtype: bfloat16
|
66 |
+
name: celeste14-20.sl
|
67 |
+
---
|
68 |
+
models:
|
69 |
+
slices:
|
70 |
+
- sources:
|
71 |
+
- layer_range: [14, 20]
|
72 |
+
model: Sao10K/L3-8B-Niitama-v1
|
73 |
+
parameters:
|
74 |
+
int8_mask: true
|
75 |
+
merge_method: passthrough
|
76 |
+
dtype: bfloat16
|
77 |
+
name: niitama14-20.sl
|
78 |
+
---
|
79 |
+
models:
|
80 |
+
- model: celeste14-20.sl
|
81 |
+
parameters:
|
82 |
+
weight: [1, 0.75, 0.625, 0.5, 0.375, 0.25, 0]
|
83 |
+
- model: niitama14-20.sl
|
84 |
+
parameters:
|
85 |
+
weight: [0, 0.25, 0.375, 0.5, 0.625, 0.75, 1]
|
86 |
+
merge_method: dare_linear
|
87 |
+
base_model: celeste14-20.sl
|
88 |
+
dtype: bfloat16
|
89 |
+
name: celeniit14-20.sl
|
90 |
+
---
|
91 |
+
models:
|
92 |
+
slices:
|
93 |
+
- sources:
|
94 |
+
- layer_range: [0, 4]
|
95 |
+
model: Nitral-AI/Hathor_Tahsin-L3-8B-v0.85
|
96 |
+
- sources:
|
97 |
+
- layer_range: [1, 5]
|
98 |
+
model: ArliAI/ArliAI-Llama-3-8B-Formax-v1.0
|
99 |
+
- sources:
|
100 |
+
- layer_range: [4, 8]
|
101 |
+
model: Nitral-AI/Hathor_Tahsin-L3-8B-v0.85
|
102 |
+
- sources:
|
103 |
+
- layer_range: [5, 9]
|
104 |
+
model: ArliAI/ArliAI-Llama-3-8B-Formax-v1.0
|
105 |
+
- sources:
|
106 |
+
- layer_range: [8, 10]
|
107 |
+
model: Sao10K/L3-8B-Niitama-v1
|
108 |
+
- sources:
|
109 |
+
- layer_range: [6, 14]
|
110 |
+
model: nothingiisreal/L3-8B-Celeste-V1.2
|
111 |
+
- sources:
|
112 |
+
- layer_range: [0, 6]
|
113 |
+
model: celeniit14-20.sl
|
114 |
+
- sources:
|
115 |
+
- layer_range: [20, 23]
|
116 |
+
model: Sao10K/L3-8B-Niitama-v1
|
117 |
+
- sources:
|
118 |
+
- layer_range: [22, 26]
|
119 |
+
model: Nitral-AI/Hathor_Tahsin-L3-8B-v0.85
|
120 |
+
- sources:
|
121 |
+
- layer_range: [22, 28]
|
122 |
+
model: nothingiisreal/L3-8B-Celeste-V1.2
|
123 |
+
- sources:
|
124 |
+
- layer_range: [25, 27]
|
125 |
+
model: Nitral-AI/Hathor_Tahsin-L3-8B-v0.85
|
126 |
+
- sources:
|
127 |
+
- layer_range: [28, 30]
|
128 |
+
model: Sao10K/L3-8B-Niitama-v1
|
129 |
+
- sources:
|
130 |
+
- layer_range: [25, 32]
|
131 |
+
model: nothingiisreal/L3-8B-Celeste-V1.2
|
132 |
+
parameters:
|
133 |
+
int8_mask: true
|
134 |
+
merge_method: passthrough
|
135 |
+
dtype: bfloat16
|
136 |
+
```
|