File size: 6,156 Bytes
9b9852d
1fbc79a
 
465b1be
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9b9852d
c9cc5f9
1fbc79a
c9cc5f9
3daa800
07b907e
1fbc79a
3a8146e
1fbc79a
1df2d19
e4d05d2
1fbc79a
 
763e27d
1fbc79a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
763e27d
 
e2a33b3
1fbc79a
 
 
 
 
 
 
 
 
 
 
 
 
465b1be
 
 
 
 
 
 
 
 
 
 
 
 
 
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
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
---
language:
- en
license: other
model-index:
- name: MythoMist-7b
  results:
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: AI2 Reasoning Challenge (25-Shot)
      type: ai2_arc
      config: ARC-Challenge
      split: test
      args:
        num_few_shot: 25
    metrics:
    - type: acc_norm
      value: 65.87
      name: normalized accuracy
    source:
      url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Gryphe/MythoMist-7b
      name: Open LLM Leaderboard
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: HellaSwag (10-Shot)
      type: hellaswag
      split: validation
      args:
        num_few_shot: 10
    metrics:
    - type: acc_norm
      value: 83.55
      name: normalized accuracy
    source:
      url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Gryphe/MythoMist-7b
      name: Open LLM Leaderboard
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: MMLU (5-Shot)
      type: cais/mmlu
      config: all
      split: test
      args:
        num_few_shot: 5
    metrics:
    - type: acc
      value: 62.32
      name: accuracy
    source:
      url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Gryphe/MythoMist-7b
      name: Open LLM Leaderboard
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: TruthfulQA (0-shot)
      type: truthful_qa
      config: multiple_choice
      split: validation
      args:
        num_few_shot: 0
    metrics:
    - type: mc2
      value: 59.98
    source:
      url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Gryphe/MythoMist-7b
      name: Open LLM Leaderboard
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: Winogrande (5-shot)
      type: winogrande
      config: winogrande_xl
      split: validation
      args:
        num_few_shot: 5
    metrics:
    - type: acc
      value: 78.06
      name: accuracy
    source:
      url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Gryphe/MythoMist-7b
      name: Open LLM Leaderboard
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: GSM8k (5-shot)
      type: gsm8k
      config: main
      split: test
      args:
        num_few_shot: 5
    metrics:
    - type: acc
      value: 20.24
      name: accuracy
    source:
      url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Gryphe/MythoMist-7b
      name: Open LLM Leaderboard
---
MythoMist 7b is, as always, a highly experimental Mistral-based merge based on my latest algorithm, which actively benchmarks the model as it's being built in pursuit of a goal set by the user.

**Addendum (2023-11-23)**: A more thorough investigation revealed a flaw in my original algorithm that has since been resolved. I've considered deleting this model as it did not follow its original objective completely but since there are plenty of folks enjoying it I'll be keeping it around. Keep a close eye [on my MergeMonster repo](https://huggingface.co/Gryphe/MergeMonster) for further developments and releases of merges produced by the Merge Monster.

The primary purpose for MythoMist was to reduce usage of the word anticipation, ministrations and other variations we've come to associate negatively with ChatGPT roleplaying data. This algorithm cannot outright ban these words, but instead strives to minimize the usage.

[The script has now been made available on my Github. Warning - Plenty of VRAM is needed.](https://github.com/Gryphe/MergeMonster/)

Quantized models are available from TheBloke: [GGUF](https://huggingface.co/TheBloke/MythoMist-7B-GGUF) - [GPTQ](https://huggingface.co/TheBloke/MythoMist-7B-GPTQ) - [AWQ](https://huggingface.co/TheBloke/MythoMist-7B-AWQ) (You're the best!)

## Final merge composition

After processing 12 models my algorithm ended up with the following (approximated) final composition:

| Model                    | Contribution |
|--------------------------|--------------|
| Neural-chat-7b-v3-1      | 26%          |
| Synatra-7B-v0.3-RP       | 22%          |
| Airoboros-m-7b-3.1.2     | 10%          |
| Toppy-M-7B               | 10%          |
| Zephyr-7b-beta           | 7%           |
| Nous-Capybara-7B-V1.9    | 5%           |
| OpenHermes-2.5-Mistral-7B| 5%           |
| Dolphin-2.2.1-mistral-7b | 4%           |
| Noromaid-7b-v0.1.1       | 4%           |
| SynthIA-7B-v1.3          | 3%           |
| Mistral-7B-v0.1          | 2%           |
| Openchat_3.5             | 2%           |

There is no real logic in how these models were divided throughout the merge - Small bits and pieces were taken from each and then mixed in with other models on a layer by layer basis, using a pattern similar to my MythoMax recipe in which underlying tensors are mixed in a criss-cross manner.

This new process only decides on the model's layers, not the singular lm_head and embed_tokens layers which influence much of the model's output. I ran a seperate script for that, picking the singular tensors that resulted in the longest responses, which settled on Toppy-M-7B.

## Prompt Format

Due to the wide variation in prompt formats used in this merge I (for now) recommend using Alpaca as the prompt template for compatibility reasons:
```
### Instruction:
Your instruction or question here.

### Response:
```

---
license: other
---
# [Open LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard)
Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/details_Gryphe__MythoMist-7b)

|             Metric              |Value|
|---------------------------------|----:|
|Avg.                             |61.67|
|AI2 Reasoning Challenge (25-Shot)|65.87|
|HellaSwag (10-Shot)              |83.55|
|MMLU (5-Shot)                    |62.32|
|TruthfulQA (0-shot)              |59.98|
|Winogrande (5-shot)              |78.06|
|GSM8k (5-shot)                   |20.24|