Datasets:
File size: 3,942 Bytes
48dfb47 88e9191 48dfb47 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 |
---
license: apache-2.0
task_categories:
- visual-question-answering
language:
- zh
pretty_name: CMMU
size_categories:
- 1K<n<10K
---
# CMMU
[**📖 Paper**](https://arxiv.org/) | [**🤗 Dataset**](https://huggingface.co/datasets) | [**GitHub**](https://github.com/FlagOpen/CMMU)
This repo contains the evaluation code for the paper [**CMMU: A Benchmark for Chinese Multi-modal Multi-type Question Understanding and Reasoning**](https://arxiv.org/) .
## Introduction
CMMU is a novel multi-modal benchmark designed to evaluate domain-specific knowledge across seven foundational subjects: math, biology, physics, chemistry, geography, politics, and history. It comprises 3603 questions, incorporating text and images, drawn from a range of Chinese exams. Spanning primary to high school levels, CMMU offers a thorough evaluation of model capabilities across different educational stages.
## Evaluation Results
We currently evaluated 10 models on CMMU. The results are shown in the following table.
| Model | Val Avg. | Test Avg. |
|----------------------------|----------|-----------|
| InstructBLIP-13b | 0.39 | 0.48 |
| CogVLM-7b | 5.55 | 4.9 |
| ShareGPT4V-7b | 7.95 | 7.63 |
| mPLUG-Owl2-7b | 8.69 | 8.58 |
| LLava-1.5-13b | 11.36 | 11.96 |
| Qwen-VL-Chat-7b | 11.71 | 12.14 |
| Intern-XComposer-7b | 18.65 | 19.07 |
| Gemini-Pro | 21.58 | 22.5 |
| Qwen-VL-Plus | 26.77 | 26.9 |
| GPT-4V | 30.19 | 30.91 |
## How to use
### Load dataset
```python
from eval.cmmu_dataset import CmmuDataset
# CmmuDataset will load *.jsonl files in data_root
dataset = CmmuDataset(data_root=your_path_to_cmmu_dataset)
```
**About fill-in-the-blank questions**
For fill-in-the-blank questions, `CmmuDataset` will generate new questions by `sub_question`, for example:
The original question is:
```python
{
"type": "fill-in-the-blank",
"question_info": "question",
"id": "subject_1234",
"sub_questions": ["sub_question_0", "sub_question_1"],
"answer": ["answer_0", "answer_1"]
}
```
Converted questions are:
```python
[
{
"type": "fill-in-the-blank",
"question_info": "question" + "sub_question_0",
"id": "subject_1234-0",
"answer": "answer_0"
},
{
"type": "fill-in-the-blank",
"question_info": "question" + "sub_question_1",
"id": "subject_1234-1",
"answer": "answer_1"
}
]
```
**About ShiftCheck**
The parameter `shift_check` is `True` by default, you can get more information about `shift_check` in our technical report.
`CmmuDataset` will generate k new questions by `shift_check`, their ids are `{original_id}-k`.
## Evaluate
The output format should be a list of json dictionaries, the required key is as follows:
```python
{
"question_id": "question id",
"answer": "answer"
}
```
Current code call gpt4 API by `AzureOpenAI`, maybe you need to modify `eval/chat_llm.py` to create your own client, and before run evaluation, you need to set environment variables like `AZURE_OPENAI_API_KEY` and `AZURE_OPENAI_ENDPOINT`.
Run
```shell
python eval/evaluate.py --result your_pred_file --data_root your_path_to_cmmu_dataset
```
**NOTE** We evaluate fill-in-the-blank questions using GPT-4 by default. If you do not have access to GPT-4, you can attempt to use a rule-based method to fill in the blanks. However, be aware that the results might differ from the official ones.
```shell
python eval/evaluate.py --result your_pred_file --data_root your_path_to_cmmu_dataset --gpt none
```
To evaluate specific type of questions, you can use `--qtype` parameter, for example:
```shell
python eval/evaluate.py --result example/gpt4v_results_val.json --data_root your_path_to_cmmu_dataset --qtype fbq mrq
```
## Citation |