alignbench如何在viz模式把结果输出到csv文件

[IcyFeather233]

之前已经跑出来了eval的结果，viz也能正常跑，但是只有这种shell里面的可视化显示，想知道怎么直接输出到csv，支持这个功能吗

另外633个任务只提取出来112个结果正常吗，gpt4的话一般是提取成功多少呢？

~/opencompass$ python run.py configs/eval_skywork_alignbench_qwen14bchat.py --debug -m viz -r 20240408_060539
04/08 10:19:20 - OpenCompass - INFO - Reusing experiements from 20240408_060539
04/08 10:19:20 - OpenCompass - DEBUG - Modules of opencompass's partitioner registry have been automatically imported from opencompass.partitioners
04/08 10:19:20 - OpenCompass - DEBUG - Get class `AlignmentBenchSummarizer` from "partitioner" registry in "opencompass"
04/08 10:19:20 - OpenCompass - DEBUG - An `AlignmentBenchSummarizer` instance is built from registry, and its implementation can be found in opencompass.summarizers.subjective.alignmentbench
Among 683 judgements, successfully extracted 112 judgements, please check!
+-----------------------+-------------------+-------------------+-------------------+--------------------+-------------------+-------------------+--------+-------------------+-------------------+
|          模型         |     事实正确性    |    满足用户需求   |       清晰度      |       完备性       |      综合得分     |     逻辑连贯性    | 创造性 |       丰富度      |  公平与可负责程度 |
+-----------------------+-------------------+-------------------+-------------------+--------------------+-------------------+-------------------+--------+-------------------+-------------------+
| qwen1.5-14b-chat-vllm | 7.767857142857143 | 7.142857142857143 | 8.714285714285714 | 7.7727272727272725 | 7.339285714285714 | 8.590909090909092 |  6.25  | 7.277777777777778 | 8.666666666666666 |
+-----------------------+-------------------+-------------------+-------------------+--------------------+-------------------+-------------------+--------+-------------------+-------------------+
+-----------------------+------------------+--------------+----------+----------+------------------+-------------------+-------------------+-------------------+-------------------+--------------------+--------------------+
|           0           |        1         |      2       |    3     |    4     |        5         |         6         |         7         |         8         |         9         |         10         |         11         |
+-----------------------+------------------+--------------+----------+----------+------------------+-------------------+-------------------+-------------------+-------------------+--------------------+--------------------+
|          模型         |       总分       |   中文推理   |          |          |     中文语言     |                   |                   |                   |                   |                    |                    |
|          模型         |       总分       | 中文推理总分 | 数学计算 | 逻辑推理 |   中文语言总分   |      基本任务     |      中文理解     |      综合问答     |      文本写作     |      角色扮演      |      专业能力      |
| qwen1.5-14b-chat-vllm | 7.32756995445156 |     7.5      |   7.0    |   8.0    | 7.15513990890312 | 7.068965517241379 | 7.588235294117647 | 6.666666666666667 | 7.142857142857143 | 6.7272727272727275 | 7.7368421052631575 |
+-----------------------+------------------+--------------+----------+----------+------------------+-------------------+-------------------+-------------------+-------------------+--------------------+--------------------+

[bittersweet1999]

1.有直接输出到csv的结果，在你的outputs下是能看到的，在summarize文件夹里
2.对于qwen14作为judge model，112/633是很正常的，它的复杂指令遵循并没有多好
3.gpt4成功基本百分比

[IcyFeather233]

3.是说的gpt4基本百分百成功吗

[bittersweet1999]

是的，基本百分百

[bittersweet1999]

如果不想付费，想用开源模型的话，推荐autoj好一点

[IcyFeather233]

Autoj好像是不是只有一个6b的是双语模型 没法用更大参数的

[bittersweet1999]

是的，你也可以尝试用https://github.com/open-compass/opencompass/blob/main/configs/datasets/subjective/alignbench/alignbench_judgeby_autoj.py
这个config，配合qwen72b试试，这个的指令更简单，应该能跟随相应的格式