--- datasets: - TIGER-Lab/MathInstruct language: - en license: apache-2.0 metrics: - accuracy pipeline_tag: text-generation --- # How Do Humans Write Code? Large Models Do It the Same Way Too Paper: [https://arxiv.org/pdf/2402.15729](https://arxiv.org/pdf/2402.15729) Code: [https://github.com/seamoke/Human-Think-Language](https://github.com/seamoke/Human-Think-Language) ## Introduction We introduce HTL, a model which utilizes the complete reasoning process of CoT to enhance PoT. This model was secondarily fine-tuned based on [MAmmoTH-Coder-7B](https://huggingface.co/TIGER-Lab/MAmmoTH-Coder-7B) ## Evaluation The models are evaluated using open-ended and multiple-choice math problems from several datasets. Here are the results: | **Model** | **GSM** |**GSM-Hard** | **NumGLUE** | **MATH** | **Sim** | **SVAMP** | **MAWPS** | **ASDiV** | |---------------------------| ----------|---------------|---------------|-----------|----------|---------- |------------|---------------| | **MAmmoTH-Coder-7B** | 59.4 |56.3 | 66.4 |33.4| 45.9 | 70.7 | 91.9 | 69.3 | | **TORA** | **72.6** |56.0 | 46.2 |**44.6**| 48.5 | 70.4 | 91.3 | **78.7** | | **MAmmoTH-Coder-7B** | 65.7 |**58.3** | **75.1** |34.9| **50.8** | **74.4** | **94.2** | 73.1 | ## Prompt Format If you want to do HTL: ``` Below is an instruction that describes a task. Write a response that appropriately completes the request. I'd like you to solve this problem in 3 steps: 1.Answer the question in plain language without writing any code.\n 2.Output one line of *\n. 3.Write program code based on the solution process in step 1 to solve the problem.\n ### Instruction: {query} Let's write a program. ### Response:" ``` ## Citation If you use the models, data, or code from this project, please cite the original paper: ``` @article{li2024humans, title={How Do Humans Write Code? Large Models Do It the Same Way Too}, author={Li, Long}, journal={arXiv preprint arXiv:2402.15729}, year={2024} } ```