File size: 3,335 Bytes
1d04e5b
 
 
 
 
 
 
 
 
 
7b1cbc7
 
5aac72a
3980f97
5aac72a
57b731d
9709984
 
 
 
94133a7
9709984
 
7b1cbc7
 
91ab63b
 
 
ea9c357
91ab63b
 
 
 
 
62d891d
57b731d
 
 
 
 
 
 
 
 
 
 
 
91ab63b
 
1d04e5b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
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
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
---
base_model:
- Sao10K/L3-8B-Stheno-v3.2
- NeverSleep/Llama-3-Lumimaid-8B-v0.1-OAS
library_name: transformers
tags:
- mergekit
- merge

---
<img src="https://huggingface.co/Alsebay/L3-8B-SMaid-v0.1/resolve/main/cover/cover.png" alt="img" style="width: 60%; min-width: 120px; height:80%; min-height: 200px; max-width:360px; max-height:600px; display: block">

> [!IMPORTANT]
> This model is wrongly using smaug-bpe pretokenizer during the merge (Mergekit uses smaug-bpe by default for llama-3 merges). So the quality of the model is worse than expected. Still usable with transformers loader.

# Update: Both version have different presents (settings) to work well
Overall:

Sao10K Stheno > SMaid V0.3 > SMaid V0.1 in Chai Benchmark

SMaid V0.1 = Sao10K Stheno > SMaid V0.3 in my custom EQ bench (Sadness and deep thought and Depression test)

Disclaimed: same seed, same character card, same scenario. 4 times try for each models.

# Best of L3-8B merge series for me. I choose 2 best variants to publish.

SMaid-V0.1: More smart, understand well content, more novelwriting. I like this version.

[SMaid-V0.3](https://huggingface.co/Alsebay/L3-8B-SMaid-v0.3): Upgrade from v0.1. More talkative, active, energetic (wrong setting, lol).

No V0.2 because I deleted it, it's a worst model of series.

I think Stheno and Lumumaid can be like a 'ying-yang', so I combine them, lol. Have test on Chaiverse, both of them got > 1995 elo score from begining. (Thanks Sao10K let me know about ChaiVerse :) )

SMaid = Stheno (it's very good) + LumiMaid (not too good, but the writing style is good)

**Recommend present (You can feedback if any setting is better)**

```
Temperature - 1.1-1.25
Min-P - 0.075
Top-K - 50
Top_P - 0.5
Repetition Penalty - 1.1
```

---

# Below is the auto-generate by Mergekit

This is a merge of pre-trained language models created using [mergekit](https://github.com/cg123/mergekit).

## Merge Details
### Merge Method

This model was merged using the [DARE](https://arxiv.org/abs/2311.03099) [TIES](https://arxiv.org/abs/2306.01708) merge method using [NeverSleep/Llama-3-Lumimaid-8B-v0.1-OAS](https://huggingface.co/NeverSleep/Llama-3-Lumimaid-8B-v0.1-OAS) as a base.

### Models Merged

The following models were included in the merge:
* [Sao10K/L3-8B-Stheno-v3.2](https://huggingface.co/Sao10K/L3-8B-Stheno-v3.2)

### Configuration

The following YAML configuration was used to produce this model:

```yaml

slices:
- sources:
  - layer_range: [0, 16]
    model: NeverSleep/Llama-3-Lumimaid-8B-v0.1-OAS
    parameters:
      density: 0.5
      weight: 1.0
  - layer_range: [0, 16]
    model: Sao10K/L3-8B-Stheno-v3.2
    parameters:
      density: 0.5
      weight: 0.9
- sources:
  - layer_range: [16, 24]
    model: Sao10K/L3-8B-Stheno-v3.2
    parameters:
      density: 0.75
      weight: 0.5
  - layer_range: [16, 24]
    model: NeverSleep/Llama-3-Lumimaid-8B-v0.1-OAS
    parameters:
      density: 0.25
      weight: 0.5
- sources:
  - layer_range: [24, 32]
    model: NeverSleep/Llama-3-Lumimaid-8B-v0.1-OAS
    parameters:
      density: 0.5
      weight: 0.5
  - layer_range: [24, 32]
    model: Sao10K/L3-8B-Stheno-v3.2
    parameters:
      density: 0.5
      weight: 1.0
merge_method: dare_ties
base_model: NeverSleep/Llama-3-Lumimaid-8B-v0.1-OAS
parameters:
int8_mask: true
dtype: bfloat16

```