New Benchmarks for Chinese LLMs

I just saw in the China AI newsletter that the new SuperCLUE Large Model benchmark evaluation from April 2024 is out. The Chinese slides are available here.

Overall, the results are in quite good agreement with our recent findings in our much more limited tests.

As you can see, several Chinese LLMs are doing quite well in the rankings. They are not in the top ranks, but they are close behind ChatGPT-4 and Claude-3 Opus. Many rank better than ChatGPT 3.5.

This is to say that, as we showed before, you can now use Chinese LLMs for most of the tasks we demonstrated in our course. Language transformation, rewriting, summarizing, programming, etc., they can all do that quite well now. The top models differ mostly in reasoning and advanced concepts, while what we are doing so far has all been rather easy for an LLM. Baidu’s model or Moonshot’s Kimi will do just fine for most tasks, as we showed.

In short: If you cannot access ChatGPT or Claude-3, the Chinese LLMs will do a good job for you and finish all the tasks we have shown so far

In the recent test, Baichuan3 ranked the best, but I am not that convinced that is is the best in China.

When asked to read and visualize a Sentinel-1 SLC image, Baichuan understands that it is a complex image and explains that in the text below the code but does not adjust the code accordingly. Still, with some follow-up questions, I am sure Baichuan3 would get there.

要使用Python读取和显示Sentinel-1合成孔径雷达（SAR）单视复数（SLC）图像，你可以使用`rasterio`库来读取图像数据，以及`matplotlib`或`pyplot`来显示图像。以下是一个简单的示例：

首先，确保你已经安装了必要的库：

```bash
pip install rasterio matplotlib
```

然后，你可以使用以下Python代码来读取并显示Sentinel-1 SLC图像：

```python
import rasterio
import matplotlib.pyplot as plt

# 替换为你的Sentinel-1 SLC文件路径
file_path = 'path/to/your/sentinel1_slc.tif'

# 使用rasterio打开SLC图像文件
with rasterio.open(file_path) as src:
    # 读取数据
    data = src.read(1)  # 假设我们只读取第一个波段

# 显示图像
plt.figure(figsize=(10, 10))
plt.imshow(data, cmap='gray')  # 使用灰度色图
plt.title('Sentinel-1 SLC Image')
plt.colorbar()  # 添加颜色条
plt.show()
```

请注意，Sentinel-1 SLC图像通常是复数值的，这意味着每个像素值包含幅度和相位信息。上面的代码仅显示了幅度信息，因为`imshow`函数通常用于可视化实数值图像。如果你需要处理复数值数据，你可能需要先对数据进行适当的转换，例如计算其幅度或相位。

此外，由于SAR图像的特性，直接显示原始数据可能不会提供很多有用的信息。通常，你会希望应用一些预处理步骤，比如多视处理、辐射定标、地理编码等，以获得更有用的信息。这些步骤可能需要更专业的SAR处理工具，如SNAP (Sentinel Application Platform) 或GDAL。

如果你需要进一步处理SAR数据，你可能需要查看专门的库，如`pyroSAR`或`sentinelsat`，它们提供了更多针对Sentinel-1数据的处理功能。