add android instruct

Author: xuyifan-0731

README.md

@@ -11,6 +11,8 @@ We develop a systematic Android agent framework，named AndroidLab. It includes
 This repository is the code framework for the operation environment and
  benchmark section. We provide two execution modes: AVD on Mac (arm64) and Docker on Linux (x86_64). You can freely add or modify new tasks or Android images according to our framework. We offer a complete evaluation framework that can be used to assess the performance of various Android agents.
 
+We have also open-sourced the Android Instruct dataset mentioned in the paper. Please refer to [here](docs/instruction_tuning.md) for more details.
+
 
 
 ![](./assets/main-picture.png)

README_CN.md

@@ -10,6 +10,7 @@
 
 该代码库是环境和测试基准的代码框架。我们提供了两种执行模式：在 Mac（arm64）上的 AVD 模式和在 Linux（x86_64）上的 Docker 模式。您可以根据我们的框架自由添加或修改新任务或 Android 镜像。我们提供了完整的评估框架，可用于评估各种 Android agents 的性能。
 
+我们也开源了文章中的Android Instruct数据集，请参考[这里](docs/instruction_tuning.md)。
 
 
 ![](./assets/main-picture.png)

docs/instruction_tuning.md

@@ -0,0 +1,9 @@
+# Android Instruct Guide
+
+## Data Download
+
+Please download the Android Instruct dataset from [this link](https://drive.google.com/file/d/1s0b74VEOww9n1kMocd6RJivwaUCymEs4/view?usp=drive_link). The dataset has been organized in the llama factory training data format. It includes 6,208 steps in XML format and 6,053 steps in SoM format. A small number of steps are missing due to certain special pages that could not be converted into the SoM format and were therefore removed. Our training is also based on this version of the dataset.
+
+## Training Details
+
+For **Llama3.1-8B**, **GLM4-9B**, **Qwen2-7B**, and **Qwen2-VL-7B**, we used the **Llama factory** framework for training. For **CogVLM** and **Llama3.2-11B-Vision**, we utilized the **Swift** framework for training. All training was conducted with a learning rate of **1e-5** and over **3 epochs**. Testing was performed using **vllm** deployment with greedy decoding.