xl-zhao/PromptCoT-Problem-Generation-Model
111
1---2license: mit3datasets:4- xl-zhao/PromptCoT-Problem-Generation-Dataset5language:6- en7base_model:8- meta-llama/Llama-3.1-8B9---10# **PromptCoT: Synthesizing Olympiad-Level Problems for Mathematical Reasoning in Large Language Modelsg** 11 12[](http://arxiv.org/abs/2503.02324) 13[](https://github.com/zhaoxlpku/PromptCoT) 14 15---16 17## 🚀 **Overview** 18The **PromptCoT Problem Generation Model** is a lightweight yet powerful model for synthesizing high-quality Olympiad-level mathematical problems. It enables the scalable construction of problem sets to facilitate post-training tasks such as **Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL)**. By systematically modeling expert problem design, PromptCoT helps generate logically consistent and intellectually demanding problems at scale.19 20For more details, refer to our **paper on ArXiv**: [🔗 PromptCoT: Synthesizing Olympiad-Level Problems for Mathematical Reasoning in Large Language Models](http://arxiv.org/abs/2503.02324). 21 22---23 24## 🔥 **Quick Start: Using the Model** 25 26### **1️⃣ Install Dependencies** 27```bash28pip install transformers vllm torch accelerate29```30 31### **2️⃣ Load the Model with Hugging Face Transformers** 32You can use the model for **direct inference** using Hugging Face’s `generate` API: 33```python34from transformers import AutoModelForCausalLM, AutoTokenizer35 36model_name = "xl-zhao/PromptCoT-Problem-Generation-Model"37tokenizer = AutoTokenizer.from_pretrained(model_name)38model = AutoModelForCausalLM.from_pretrained(model_name).to("cuda")39 40foundational_concepts = [41 "Ability to apply quantitative reasoning and estimation techniques to solve problems, including making approximations and using logical deductions to arrive at a solution.",42 "Ability to solve equations involving complex numbers, including finding conditions under which two complex numbers are equal, particularly in the context of their magnitudes and arguments.",43 "Fractional arithmetic: Performing calculations with fractions to determine the final probability.",44 "Interpreting and solving problems involving nested operations or functions.",45 "Using logical reasoning to connect given data points and derive conclusions."46]47 48difficulty_level = "HMMT-Feb"49 50prompt = (51 "Given foundational concepts and difficulty level, identify connections and develop a question "52 "that integrates these concepts with appropriate complexity.\n\n"53 "Foundational Concepts:\n"54 + "\n".join(f"{i+1}. {concept}" for i, concept in enumerate(foundational_concepts))55 + f"\n\nDifficulty Level: {difficulty_level}"56)57 58inputs = tokenizer(prompt, return_tensors="pt").to("cuda")59 60with torch.no_grad():61 output = model.generate(**inputs, max_length=4096, temperature=0.6)62 63generated_problem = tokenizer.decode(output[0], skip_special_tokens=True)64print(generated_problem)65```66 67---68 69## ⚡ **Using vLLM for Fast Inference** 70For optimized inference, use `vLLM`: 71```python72from vllm import LLM, SamplingParams73 74model_name = "xl-zhao/PromptCoT-Problem-Generation-Model"75llm = LLM(model=model_name, tensor_parallel_size=1)76 77foundational_concepts = [78 "Ability to apply quantitative reasoning and estimation techniques to solve problems, including making approximations and using logical deductions to arrive at a solution.",79 "Ability to solve equations involving complex numbers, including finding conditions under which two complex numbers are equal, particularly in the context of their magnitudes and arguments.",80 "Fractional arithmetic: Performing calculations with fractions to determine the final probability.",81 "Interpreting and solving problems involving nested operations or functions.",82 "Using logical reasoning to connect given data points and derive conclusions."83]84 85difficulty_level = "HMMT-Feb"86 87prompt = (88 "Given foundational concepts and difficulty level, identify connections and develop a question "89 "that integrates these concepts with appropriate complexity.\n\n"90 "Foundational Concepts:\n"91 + "\n".join(f"{i+1}. {concept}" for i, concept in enumerate(foundational_concepts))92 + f"\n\nDifficulty Level: {difficulty_level}"93)94 95sampling_params = SamplingParams(temperature=0.6, max_tokens=4096)96outputs = llm.generate([prompt], sampling_params)97 98print(outputs[0].outputs[0].text)99```100 101---102 103## 🔗 **Full Usage & Advanced Options** 104For advanced usage, including **batch inference and rejection sampling for filtering high-quality problems**, refer to the **full repository on GitHub**: 105🔹 [GitHub: PromptCoT](https://github.com/zhaoxlpku/PromptCoT) 106 107---108 109## 📜 **Citation** 110If you use **PromptCoT**, please consider citing: 111```112@article{zhao2025promptcot,113 author = {Zhao, Xueliang and Wu, Wei and Guan, Jian and Kong, Lingpeng},114 title = {PromptCoT: Synthesizing Olympiad-Level Problems for Mathematical Reasoning in Large Language Models},115 year = {2025},116 journal = {arXiv preprint arXiv:2503.02324},117 url = {http://arxiv.org/abs/2503.02324}118}119```120 121 122 123