gemma3n-slicing-configs

This repository contains configurations to slice Gemma 3n E4B, which is enabled thanks to it being a MatFormer.
The E4B model can be sliced into small models, trading off quality and latency/compute requirements.
We recommend exploring the [MatFormer Lab](TODO: add link) to getting started with slicing Gemma 3n E4B yourself.

For each configuration, we calculate the MMLU accuracy.
Although these are not the only configurations possible, they are optimal configurations
identified by calculating the accuracy of the pre-trained model

To learn more about MatFormers, please review the and generate your own submodels
with the [MatFormer Lab](TODO: add link).

This chart show’s MMLU performance vs model size of Gemma 3n Mix-n-Match (pretrained) capability.

Some additional resources:

* [Gemma 3n launch blog](https://developers.googleblog.com/en/introducing-gemma-3n-developer-guide)
* [MatFormer paper](https://huggingface.co/papers/2310.07707)

本仓库包含Gemma 3n E4B的切片配置，该模型凭借其MatFormer架构得以支持切片操作。E4B模型可被切片为多个小型模型，以此在模型质量与延迟/计算资源需求之间进行权衡。我们建议您访问[MatFormer实验室（MatFormer Lab）](TODO: add link)，以自行上手开展Gemma 3n E4B的切片工作。

针对每一种配置，我们均计算了其MMLU准确率。尽管此类配置并非全部可行方案，但它们均为通过测算预训练模型准确率所得到的最优配置。

若欲深入了解MatFormer架构，请查阅相关资料，并通过[MatFormer实验室（MatFormer Lab）](TODO: add link)生成自定义子模型。

本图表展示了Gemma 3n Mix-n-Match（预训练）模型的MMLU性能与模型尺寸的对应关系。

其他相关资源：
* [Gemma 3n发布博客](https://developers.googleblog.com/en/introducing-gemma-3n-developer-guide)
* [MatFormer研究论文](https://huggingface.co/papers/2310.07707)