Z1 7B

🚀 Z1: 基于代码的高效测试时缩放

本项目旨在训练大语言模型，使其能够以转换思维的方式进行推理，为相关领域的研究和应用提供了新的思路和方法。

🚀 快速开始

若要开启转换思维模式，请参考 https://github.com/efficientscaling/Z1。

✨ 主要特性

• 基础模型：Qwen/Qwen2.5-Coder-7B-Instruct
• 库名称：transformers
• 许可证：MIT
• 评估指标：准确率
• 任务类型：文本生成

📚 详细文档

论文

• 📜 Paper

模型

• 🤗 HF Models

代码仓库

• 🐱 GitHub

💻 使用示例

基础用法

import copy
from typing import List
from dataclasses import dataclass

import gradio as gr
from vllm import LLM, SamplingParams
from transformers import  AutoTokenizer

BOX=r"\boxed{}"
ANSWER_WITH_BOX=f"\n\nI overthought it, the final answer in {BOX} should be:\n\n"
ANSWER_WITHOUT_BOX=f"\n\nI overthought it, the final answer should be:\n\n"

model_name = "efficientscaling/Z1-7B"

@dataclass
class ThinkingLLM(LLM):

    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)

    def thinking_generate(self, prompts: List[str], sampling_params: SamplingParams = None, max_tokens_for_thinking: int = None):

        # If no SamplingParams is provided, create a default one
        if sampling_params is None:
            raise ValueError("Sampling_params can't be None!")
        else:
            all_max_tokens = sampling_params.max_tokens
            # Override the max_tokens in the provided SamplingParams with the budget
            sampling_params.max_tokens = max_tokens_for_thinking
            print(f"All tokens: {all_max_tokens}")
            print(f"Tokens for thinking: {max_tokens_for_thinking}")

        trajectories = self.generate(prompts, sampling_params)

        rethinking_str = ANSWER_WITHOUT_BOX
        sampling_params.max_tokens = all_max_tokens

        answers = copy.deepcopy(trajectories)

        unfinished_id = []
        thinking_token = 0
        new_prompts = []

        for id, traj in enumerate(trajectories):
            if traj.outputs[0].finish_reason == 'length':
                unfinished_id.append(id)
                new_prompts.append(prompts[id] + traj.outputs[0].text + rethinking_str)
            thinking_token += len(traj.outputs[0].token_ids)

        avg_thinking_token = thinking_token / len(prompts)

        if new_prompts:
            print(new_prompts[0])

            o = self.generate(
                new_prompts,
                sampling_params=sampling_params,
            )
            
        for i, uid in enumerate(unfinished_id):
            answers[uid] = o[i]

        return new_prompts, answers


def generate_text(prompt, max_tokens, max_tokens_for_thinking, temperature, top_p):

    sampling_params = SamplingParams(
        temperature=temperature,
        max_tokens=max_tokens,
        top_p=top_p,
        skip_special_tokens=False,
    )

    trajectories, outputs = llm.thinking_generate(prompt, sampling_params, max_tokens_for_thinking=max_tokens_for_thinking)
    return trajectories[0] + '\n\n' + outputs[0].outputs[0].text if trajectories else outputs[0].outputs[0].text


llm = ThinkingLLM(
    model=model_name,
    tensor_parallel_size=1,
    gpu_memory_utilization=0.96,
    )


with gr.Blocks() as demo:
    gr.Markdown("# Reason with shifted thinking")
    
    with gr.Row():
        with gr.Column():
            prompt_input = gr.Textbox(
                label="Prompt",
                placeholder="Input",
                lines=5,
            )
            max_tokens_for_thinking_input = gr.Slider(
                label="shifted_thinking_window_size",
                minimum=1,
                maximum=32786,
                value=4000,
                step=1,
            )
            max_tokens_input = gr.Slider(
                label="all_max_tokens",
                minimum=1,
                maximum=32786,
                value=32786,
                step=1,
            )
            temperature_input = gr.Slider(
                label="Temperature",
                minimum=00,
                maximum=2.0,
                value=0,
                step=0.1,
            )
            top_p_input = gr.Slider(
                label="Top-p",
                minimum=0.0,
                maximum=1.0,
                value=1,
                step=0.01,
            )
            generate_button = gr.Button("Generate")
        
        with gr.Column():
            output_text = gr.Textbox(
                label="Shifted Thinking Window",
                placeholder="Text is here...",
                lines=10,
            )
    
    generate_button.click(
        fn=generate_text,
        inputs=[prompt_input, max_tokens_for_thinking_input,max_tokens_input,  temperature_input, top_p_input],
        outputs=output_text, 
    )

if __name__ == "__main__":
    demo.launch()

📄 许可证

本项目采用 MIT 许可证。

📜 引用

@misc{yu2025efficientscaling,
      title={Z1: Efficient Test-time Scaling with Code}, 
      author={Zhaojian Yu and Yinghao Wu and Yilun Zhao and Arman Cohan and Xiao-Ping Zhang},
      year={2025},
      eprint={2504.00810},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2504.00810}, 
}