MathEval

MathEval is a benchmark dedicated to a comprehensive evaluation of the mathematical capabilities of large models. It encompasses over 20 evaluation datasets across various mathematical domains, with over 30,000 math problems. The goal is to thoroughly evaluate the performance of large models in tackling problems spanning a wide range of difficulty levels and diverse mathematical subfields (i.e. arithmetic, elementary mathematics, middle and high school competition topics, advanced mathematical, etc.). It serves as the trustworthy reference for cross-model comparisons of mathematical abilities among large models at the current stage and provides guidance on how to further enhance the mathematical capabilities of these models in the future.

Evaluation Pipeline

The whole pipeline is time consuming, we recommend to submit your evaluation request through our website

Step0: Download Dataset

Download the evaluation datasets from Google Drive

Step1: Dataset preprocessing

We unify the input format of different mathematical dataset as following:

For zero shot settings:

{"conversations": [{"from":"human", "value":"xxx"}]}

For few shot settings:

{"conversations": [{"from":"human", "value":"example0-question"}, {"from":"gpt", "value":"example0-answer"}, {},{}, ...]}

Running Script:

python preprocess_dataset/build_dataset.py --input_dir ./datasets --output_dir ./output

where ./datasets is the path where you save your downloaded datasets

Step2: How to construct the input prompt and the inference script

We format the input prompt for different base model and their chat edition based on the fastchat.

• If you have elaborate templates for your models, implement your templates inside the conversation.py file, and change the settings in generate_shell_config.py file with your own template_name;
• You can also use default template_name for your model. Please also change the data_dir variable in generate_shell_config.py file. The generate_shell_config.py contains all the preprocessing config for each dataset and each base model.
• We assume you have at least 2 80G-GPUs in your device to run over 70B size models. You need to ensure that the model can be successfully loaded with the given num_gpus. Please run the following script to generate the inference scripts, you need to specify the path to the model's output directory, the path to the log directory, and the path to save the running shell scripts:

Step3: How to Run your model with the processed dataset

Using VLLM for Inference Acceleration

We now only support using VLLM for inference acceleration. To enable VLLM acceleration, use the following command:

python run.py --configs ./generate_shell_config.py --accelerator vllm

Installing VLLM

To use VLLM for inference acceleration, you need to install it first. Please refer to the VLLM GitHub repository for installation instructions.

Running Mode

We now support two modes: "infer" and "eval". In "infer" mode, the system only generates model outputs. In "eval" mode, the system not only generates model outputs but also compares them with the real answers using the compare-model.

The default mode is "infer". To start in "eval" mode, you can use the following command:

python run.py --configs ./generate_shell_config.py --accelerator vllm --mode eval

We have separated the modes for greater flexibility. You can compare the answers using either GPT-4 or the compare-model.

Parameter Explanation

• --configs ./generate_shell_config.py: Specifies the path to the configuration file.
• --accelerator vllm: Enables VLLM for inference acceleration.

Step 4: Compare-Answer

If you have access to GPT4, please jump to Step 5.

If you do not have access to GPT4, we have released a compare-answer model in HuggingFace, please refer to: DeepSeek-Math-Compare-Answer

Now we implement compare compare_with_local_model.py

Example Command

python compare_with_local_model.py --model_path /path/to/model --input_dir /path/to/input --output_dir /path/to/output --device_num 4

Command Line Arguments

• --model_path: Path to the pre-trained model directory.
• --input_dir: Directory containing the input JSON files.
• --output_dir: Directory to save the output JSON files.
• --device_num: Number of GPUs to use for parallel processing.

Step5: Answer-compare with GPT4

How to extract the answer

We have provide our prompt in ./prompts/extraction_prompts folder.

We recommend to realize the send_chat_request function in run_gpt4_extraction.py by yourself, since we do not know the detailed GPT4 invoking function of yours.

How to verify the answer generated from GPT4 and the answer from the golden dataset

We have provide our prompt in ./prompts/verification_prompts folder.

We recommend to realize the send_chat_request function in run_gpt4_extraction.py by yourself, since we do not know the detailed GPT4 invoking function of yours.

Discussion

Why we want to use GPT4 for answer extraction and answer verification.

precision comparison

The precision comparison of using GPT4 and REGEX from OpenCompass for answer extractor

Corner cases for extraction

The precision comparison of using GPT4 and REGEX from OpenCompass for answer verification

Our total cost

Dataset