File size: 7,999 Bytes
4300638 6514a17 4300638 6514a17 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 |
---
library_name: transformers
language:
- ko
base_model:
- meta-llama/Llama-3.1-8B
- NCSOFT/Llama-VARCO-8B-Instruct
- akjindal53244/Llama-3.1-Storm-8B
pipeline_tag: text-generation
---
# ๐ค LLM Evolutionary Merge
๐ค [Model](https://huggingface.co/fiveflow/LLMEvoLLaMA-3.1-8B-v0.1) | ๐ [Github](https://github.com/kwon13/LLM-Evo-Merge) | โ๏ธ [Blog](์์ฑ์ค..) | ๐ก[Inspired by Sakana AI](https://github.com/SakanaAI/evolutionary-model-merge)
This project aims to optimize model merging by integrating LLMs into evolutionary strategies in a novel way. Instead of using the [CMA-ES](https://en.wikipedia.org/wiki/CMA-ES) approach, the goal is to improve model optimization by [leveraging the search capabilities of LLMs](https://arxiv.org/abs/2402.18381) to explore the parameter space more efficiently and adjust the search scope based on high-performing solutions.
Currently, the project supports optimization only within the Parameter Space, but I plan to extend its functionality to enable merging and optimization in the Data Flow Space as well. This will further enhance model merging by optimizing the interaction between data flow and parameters.
## Performance
I focused on creating a high-performing Korean model solely through merging, without additional model training.
<details>
<summary>Merging Recipe</summary>
```YAML
base_model: meta-llama/Llama-3.1-8B
dtype: bfloat16
merge_method: task_arithmetic
allow_negative_weights: true
parameters:
int8_mask: 1.0
normalize: 1.0
slices:
- sources:
- layer_range: [0, 2]
model: NCSOFT/Llama-VARCO-8B-Instruct
parameters:
weight: 1
- layer_range: [0, 2]
model: akjindal53244/Llama-3.1-Storm-8B
parameters:
weight: 0.3475802891062396
- layer_range: [0, 2]
model: meta-llama/Llama-3.1-8B
- sources:
- layer_range: [2, 4]
model: NCSOFT/Llama-VARCO-8B-Instruct
parameters:
weight: 0.8971381657317269
- layer_range: [2, 4]
model: akjindal53244/Llama-3.1-Storm-8B
parameters:
weight: 0.45369921781118544
- layer_range: [2, 4]
model: meta-llama/Llama-3.1-8B
- sources:
- layer_range: [4, 6]
model: NCSOFT/Llama-VARCO-8B-Instruct
parameters:
weight: 0.5430828084884667
- layer_range: [4, 6]
model: akjindal53244/Llama-3.1-Storm-8B
parameters:
weight: 0.2834723715836387
- layer_range: [4, 6]
model: meta-llama/Llama-3.1-8B
- sources:
- layer_range: [6, 8]
model: NCSOFT/Llama-VARCO-8B-Instruct
parameters:
weight: 0.419043948030593
- layer_range: [6, 8]
model: akjindal53244/Llama-3.1-Storm-8B
parameters:
weight: 0.3705268601566145
- layer_range: [6, 8]
model: meta-llama/Llama-3.1-8B
- sources:
- layer_range: [8, 10]
model: NCSOFT/Llama-VARCO-8B-Instruct
parameters:
weight: 0.3813333860404775
- layer_range: [8, 10]
model: akjindal53244/Llama-3.1-Storm-8B
parameters:
weight: 0.7634501436288518
- layer_range: [8, 10]
model: meta-llama/Llama-3.1-8B
- sources:
- layer_range: [10, 12]
model: NCSOFT/Llama-VARCO-8B-Instruct
parameters:
weight: 0.49134830660275863
- layer_range: [10, 12]
model: akjindal53244/Llama-3.1-Storm-8B
parameters:
weight: 0.7211994938499454
- layer_range: [10, 12]
model: meta-llama/Llama-3.1-8B
- sources:
- layer_range: [12, 14]
model: NCSOFT/Llama-VARCO-8B-Instruct
parameters:
weight: 0.9218963071448836
- layer_range: [12, 14]
model: akjindal53244/Llama-3.1-Storm-8B
parameters:
weight: 0.5117022419864319
- layer_range: [12, 14]
model: meta-llama/Llama-3.1-8B
- sources:
- layer_range: [14, 16]
model: NCSOFT/Llama-VARCO-8B-Instruct
parameters:
weight: 0.8238938467581831
- layer_range: [14, 16]
model: akjindal53244/Llama-3.1-Storm-8B
parameters:
weight: 0.851712316016478
- layer_range: [14, 16]
model: meta-llama/Llama-3.1-8B
- sources:
- layer_range: [16, 18]
model: NCSOFT/Llama-VARCO-8B-Instruct
parameters:
weight: 0.3543028846914006
- layer_range: [16, 18]
model: akjindal53244/Llama-3.1-Storm-8B
parameters:
weight: 0.6864368345788241
- layer_range: [16, 18]
model: meta-llama/Llama-3.1-8B
- sources:
- layer_range: [18, 20]
model: NCSOFT/Llama-VARCO-8B-Instruct
parameters:
weight: 0.9189961100847883
- layer_range: [18, 20]
model: akjindal53244/Llama-3.1-Storm-8B
parameters:
weight: 0.5800251781306379
- layer_range: [18, 20]
model: meta-llama/Llama-3.1-8B
- sources:
- layer_range: [20, 22]
model: NCSOFT/Llama-VARCO-8B-Instruct
parameters:
weight: 0.9281691677008521
- layer_range: [20, 22]
model: akjindal53244/Llama-3.1-Storm-8B
parameters:
weight: 0.5356892784211416
- layer_range: [20, 22]
model: meta-llama/Llama-3.1-8B
- sources:
- layer_range: [22, 24]
model: NCSOFT/Llama-VARCO-8B-Instruct
parameters:
weight: 0.839268407952539
- layer_range: [22, 24]
model: akjindal53244/Llama-3.1-Storm-8B
parameters:
weight: 0.5082186376599986
- layer_range: [22, 24]
model: meta-llama/Llama-3.1-8B
- sources:
- layer_range: [24, 26]
model: NCSOFT/Llama-VARCO-8B-Instruct
parameters:
weight: 0.6241902192095534
- layer_range: [24, 26]
model: akjindal53244/Llama-3.1-Storm-8B
parameters:
weight: 0.2945221540685877
- layer_range: [24, 26]
model: meta-llama/Llama-3.1-8B
- sources:
- layer_range: [26, 28]
model: NCSOFT/Llama-VARCO-8B-Instruct
parameters:
weight: 0.7030728026501202
- layer_range: [26, 28]
model: akjindal53244/Llama-3.1-Storm-8B
parameters:
weight: 0.2350478509634181
- layer_range: [26, 28]
model: meta-llama/Llama-3.1-8B
- sources:
- layer_range: [28, 30]
model: NCSOFT/Llama-VARCO-8B-Instruct
parameters:
weight: 0.2590342230366074
- layer_range: [28, 30]
model: akjindal53244/Llama-3.1-Storm-8B
parameters:
weight: 0.006083182855312869
- layer_range: [28, 30]
model: meta-llama/Llama-3.1-8B
- sources:
- layer_range: [30, 32]
model: NCSOFT/Llama-VARCO-8B-Instruct
parameters:
weight: 1
- layer_range: [30, 32]
model: akjindal53244/Llama-3.1-Storm-8B
parameters:
weight: 0.234650395825126
- layer_range: [30, 32]
model: meta-llama/Llama-3.1-8B
```
</details>
The models used for merging are listed below.
```
Base Model: meta-llama/Llama-3.1-8B
Model 1: NCSOFT/Llama-VARCO-8B-Instruct
Model 2: akjindal53244/Llama-3.1-Storm-8B
```
### Comparing LLMEvoLlama with Source in Korean Benchmark
- LogicKor: A benchmark that evaluates various linguistic abilities in Korean, including math, writing, coding, comprehension, grammar, and reasoning skills. (https://lk.instruct.kr/)
- KoBest: A benchmark consisting of five natural language understanding tasks designed to test advanced Korean language comprehension. (https://arxiv.org/abs/2204.04541)
### Comparing LLMEvoLlama with Source in English Benchmark and Total Average
| Model | truthfulqa_mc2 (0-shot acc) | arc_challenge (0-shot acc) | Korean + English Performance (avg) |
|-----------------|-------------------------|------------------------|------------------------------|
| [VARCO](https://huggingface.co/NCSOFT/Llama-VARCO-8B-Instruct) | 0.53 | 0.47 | 0.68 |
| [Llama-Instruct](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct) | 0.53 | 0.52 | 0.66 |
| [Llama-Storm](https://huggingface.co/akjindal53244/Llama-3.1-Storm-8B) | 0.59 | 0.52 | 0.67 |
| [LLMEvoLLaMA](https://huggingface.co/fiveflow/LLMEvoLLaMA-3.1-8B-v0.1) | 0.57 | 0.50 | **0.71** |
|