### Install Project Dependencies

Source: https://github.com/sakanaai/evolutionary-model-merge/blob/main/README.md

Install the required Python packages for the project using pip. This command installs the project in editable mode. The recommended environment is Python 3.10.12 and CUDA 12.3.

```bash
pip install -e .
```

--------------------------------

### Clone the Project Repository

Source: https://github.com/sakanaai/evolutionary-model-merge/blob/main/README.md

Clone the evolutionary-model-merge repository from GitHub and change the current directory to the newly created project folder to proceed with the setup.

```bash
git clone https://github.com/SakanaAI/evolutionary-model-merge.git
cd evolutionary-model-merge
```

--------------------------------

### Run Evaluation Script

Source: https://github.com/sakanaai/evolutionary-model-merge/blob/main/README.md

Launch the evaluation process by running the `evaluate.py` script. You must specify the path to a configuration file, which can be found in the `configs` directory.

```bash
python evaluate.py --config_path {path-to-config}
```

--------------------------------

### Benchmark: EvoLLM-JP vs. Source LLMs

Source: https://github.com/sakanaai/evolutionary-model-merge/blob/main/README.md

Performance comparison of EvoLLM-JP models against their source Large Language Models (LLMs) on the MGSM-JA and lm-eval-harness benchmarks. Higher scores indicate better performance.

```APIDOC
Benchmark: MGSM-JA (acc ↑) | lm-eval-harness (avg ↑)
--------------------------------------------------------
Model: Shisa Gamma 7B v1
Scores: 9.6 | 66.1

Model: WizardMath 7B V1.1
Scores: 18.4 | 60.1

Model: Abel 7B 002
Scores: 30.0 | 56.5

Model: Arithmo2 Mistral 7B
Scores: 24.0 | 56.4

Model: EvoLLM-JP-A-v1-7B
Scores: 52.4 | 69.0

Model: EvoLLM-JP-v1-7B
Scores: 52.0 | 70.5

Model: EvoLLM-JP-v1-10B
Scores: 55.6 | 66.2
```

--------------------------------

### Released Evolutionary Merged Models

Source: https://github.com/sakanaai/evolutionary-model-merge/blob/main/README.md

A list of models developed by SakanaAI using the Evolutionary Model Merge technique. The table includes model names, size, license type, and the source models used for the merge.

```APIDOC
Model: EvoLLM-JP-v1-7B
Size: 7B
License: Microsoft Research License
Source: shisa-gamma-7b-v1, WizardMath-7B-V1.1, GAIR/Abel-7B-002

Model: EvoLLM-JP-v1-10B
Size: 10B
License: Microsoft Research License
Source: EvoLLM-JP-v1-7B, shisa-gamma-7b-v1

Model: EvoLLM-JP-A-v1-7B
Size: 7B
License: Apache 2.0
Source: shisa-gamma-7b-v1, Arithmo2-Mistral-7B, GAIR/Abel-7B-002

Model: EvoVLM-JP-v1-7B
Size: 7B
License: Apache 2.0
Source: LLaVA-1.6-Mistral-7B, shisa-gamma-7b-v1
```

--------------------------------

### Benchmark: EvoVLM-JP vs. Existing VLMs

Source: https://github.com/sakanaai/evolutionary-model-merge/blob/main/README.md

Performance comparison of the EvoVLM-JP model against other existing Vision-Language Models (VLMs) on Japanese benchmarks. Higher ROUGE-L scores indicate better performance. Note: Japanese Stable VLM was not evaluated on JA-VG-VQA-500 as it was part of its training data.

```APIDOC
Benchmark: JA-VG-VQA-500 (ROUGE-L ↑) | JA-VLM-Bench-In-the-Wild (ROUGE-L ↑)
--------------------------------------------------------------------------------
Model: LLaVA-1.6-Mistral-7B
Scores: 14.32 | 41.10

Model: Japanese Stable VLM
Scores: -*1 | 40.50

Model: Heron BLIP Japanese StableLM Base 7B llava-620k
Scores: 14.51 | 33.26

Model: EvoVLM-JP-v1-7B
Scores: 19.70 | 51.25

*1: Japanese Stable VLM cannot be evaluated using the JA-VG-VQA-500 dataset because this model has used this dataset for training.
```

=== COMPLETE CONTENT === This response contains all available snippets from this library. No additional content exists. Do not make further requests.