🐟 Evolutionary Optimization of Model Merging Recipes

🤗 Models | 👀 Demo | 📚 Paper | 📝 Blog | 🐦 Twitter

This repository serves as a central hub for SakanaAI's Evolutionary Model Merge series, showcasing its releases and resources. It includes models and code for reproducing the evaluation presented in our paper. Look forward to more updates and additions coming soon.

Models

Our Models

Model	Size	License	Source
EvoLLM-JP-v1-7B	7B	Microsoft Research License	shisa-gamma-7b-v1, WizardMath-7B-V1.1, GAIR/Abel-7B-002
EvoLLM-JP-v1-10B	10B	Microsoft Research License	EvoLLM-JP-v1-7B, shisa-gamma-7b-v1
EvoLLM-JP-A-v1-7B	7B	Apache 2.0	shisa-gamma-7b-v1, Arithmo2-Mistral-7B, GAIR/Abel-7B-002
EvoVLM-JP-v1-7B	7B	Apache 2.0	LLaVA-1.6-Mistral-7B, shisa-gamma-7b-v1

Comparing EvoLLM-JP w/ Source LLMs

For details on the evaluation, please refer to Section 4.1 of the paper.

Model	MGSM-JA (acc ↑)	lm-eval-harness (avg ↑)
Shisa Gamma 7B v1	9.6	66.1
WizardMath 7B V1.1	18.4	60.1
Abel 7B 002	30.0	56.5
Arithmo2 Mistral 7B	24.0	56.4
EvoLLM-JP-A-v1-7B	52.4	69.0
EvoLLM-JP-v1-7B	52.0	70.5
EvoLLM-JP-v1-10B	55.6	66.2

Comparing EvoVLM-JP w/ Existing VLMs

For details on the evaluation, please see Section 4.2 of the paper.

Model	JA-VG-VQA-500 (ROUGE-L ↑)	JA-VLM-Bench-In-the-Wild (ROUGE-L ↑)
LLaVA-1.6-Mistral-7B	14.32	41.10
Japanese Stable VLM	-^*1	40.50
Heron BLIP Japanese StableLM Base 7B llava-620k	8.73^*2	27.37^*2
EvoVLM-JP-v1-7B	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.
*2: We are checking with the authors to see if this current results are valid.

Reproducing the Evaluation

1. Clone the Repo

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

2. Download fastext Model

We use fastext to detect language for evaluation. Please download lid.176.ftz from this link and place it in your current directory. If you place the file in a directory other than the current directory, specify the path to the file using the LID176FTZ_PATH environment variable.

3. Install Libraries

pip install -e .

We conducted our tests in the following environment: Python Version 3.10.12 and CUDA Version 12.3. We cannot guarantee that it will work in other environments.

4. Run

To launch evaluation, run the following script with a certain config. All configs used for the paper are in configs.

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

Acknowledgement

We would like to thank the developers of the source models for their contributions and for making their work available. Our math evaluation code builds on the WizardMath repository, and we are grateful for their work.

switiz/evolutionary-model-merge