Paper Link This repo introduces Xwin-Math, a series of powerful SFT LLMs for math problem based on LLaMA-2. Moreover, it also includes an evaluation toolkit that better converts LaTeX formulas into SymPy objects, enabling more accurate assessment of the mathematical abilities.

• 💥 [May, 2024] The Xwin-Math-70B-V1.1 model achieves 51.9 pass@1 on the MATH benchmark and 90.6 pass@1 on the GSM8K benchmark. This is a new SoTA model based on LLaMA-2-70B!
• 💥 [May, 2024] The Xwin-Math-7B-V1.1 model achieves 44.7 pass@1 on the MATH benchmark and 84.4 pass@1 on the GSM8K benchmark. This is a new SoTA model based on LLaMA-2-7B!
• 💥 [Nov, 2023] The Xwin-Math-70B-V1.0 model achieves 31.8 pass@1 on the MATH benchmark and 87.0 pass@1 on the GSM8K benchmark. This performance places it first amongst all open-source models!
• 💥 [Nov, 2023] The Xwin-Math-7B-V1.0 and Xwin-Math-13B-V1.0 models achieve 66.6 and 76.2 pass@1 on the GSM8K benchmark, ranking as top-1 among all LLaMA-2 based 7B and 13B open-source models, respectively!

• Xwin-Math-7B-V1.1 uses 1.92M GSM8K and 960K MATH synthetic data.
• Xwin-Math-70B-V1.1 uses 960K GSM8K and 480K MATH synthetic data.

Xwin-Math-70B-V1.0 has achieved 31.8% on MATH and 87.0% on GSM8K. These scores are 5.3 and 3.1 points higher, respectively, than the previous state-of-the-art open-source MetaMath and LEMAv1 model.

We obtain these results using our flexible evaluation strategy. Due to differences in environment and hardware, the numbers may be different from the reported results, but we ensure that the evaluation is as accurate and fair as possible.

Our 70B model shows strong mathematical synthesis capabilities among all open-sourced models. Also note that our model even approaches or surpasses the performance of GPT-35-Turbo on some benchmarks.

* LEMAv1 has two models, and we report the better LEMAv1-70B-MATH model in these benchmarks.

In order to evaluate a model's mathematical capabilities more flexibly and ensure a fair comparison of results, particularly for the MATH benchmark, we have developed a new evaluation tool. We have also assessed the pass@1 results of recent models on MATH and GSM8K benchmarks, which provides more accurate results.

We hope this toolkit can benefit open-source community by providing more accurate insights and conclusions. For a deeper understanding of our evaluation tool and methods, please visit here

• "Report" refers to the accuracy stated in the original papers.
• "Repro" indicates the results is reproduced by generating responses and evaluating them using the respective open-source models and scripts.
• "Strict" and "Flex" denote the results we achieved by employing our two answer extraction strategies and evaluate the same responses as "Repro".

Before you start, please install the requirements.

pip install -r requirements.txt

We tested our result using python 3.8 and cuda 11.8. We recommend you use docker.

docker run --gpus all -it --rm --ipc=host superbench/dev:cuda11.8

To generate the model's responses, you can use the generate.py script. Please be aware that generating responses is separate from verifying their correctness. After that, we will then check for their correctness.

For the generation process, we use the Vicuna-v1.1 system prompt with chain-of-thought and format instruction. We also employ a greedy decoding strategy and set the maximum sequence length to 2048.

"A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: {instruction} Give your solution in detail. In the end, write your final answer in the format of 'The answer is: <ANSWER>.'. ASSISTANT:"

Here is an simple example to generate using vLLM.

cd eval

python generate.py --dataset_path dataset/gsm8k.json --model_path path/to/your/model --tensor_parallel_size 4

By default the results will be output to the eval/response, using the prompt eval/prompt/xwin_math.json. If you wish to change the output path or use a different prompt

python generate.py --dataset_path dataset/gsm8k.json --model_path path/to/your/model --tensor_parallel_size 4 --output_path /your/path --prompt_path /your/path

We provide some datasets (in eval/dataset):

• gsm8k.json: GSM8K.
• math.json: MATH.
• combination.json: A combination of many benchmarks, can evaluate the OOD capability of the model.

If you wan't to use your own datasets, please format your dataset like this.

[
    {
        "question": "Janet\u2019s ducks lay 16 eggs per day. She eats three for breakfast every morning and bakes muffins for her friends every day with four. She sells the remainder at the farmers' market daily for $2 per fresh duck egg. How much in dollars does she make every day at the farmers' market?",
        "answer": "18",
        "type": "GSM8K",
        "subtype": "",
        "level": 0,
    },
    // ... more data items
]

To verify the accuracy of the answers after generation, you can use the `check.py script.

Here is an simple example

cd eval

python eval.py /path/to/model/response

The result will be saved in eval/evaluation

If you do not want to save the results or want to change the save path

python eval.py --data_path /path/to/model/response --save_path /path/to/save --save_result True

Once you run the script, the terminal will display the output as a table. This table will show the number of instances for each benchmark and the corresponding accuracy. Here is a hypothetical example of what the output might look like:

Please consider citing our work if you use the data or code in this repo.

@software{xwin-math,
  title = {Xwin-Math},
  author = {Xwin-Math Team},
  url = {https://github.com/Xwin-LM/Xwin-LM/Xwin-Math},
  version = {pre-release},
  year = {2023},
  month = {11},
}

Thanks to Llama 2, FastChat, and vLLM.