bespoke-manim

<a href="https://github.com/bespokelabsai/curator/">
<img src="https://huggingface.co/datasets/bespokelabs/Bespoke-Stratos-17k/resolve/main/made_with_curator.png" alt="Made with Curator" width=200px>
</a>

The Bespoke-Manim is a fully synthetic collection of 1,000 math animation video scripts, accompanied by corresponding Manim-generated animations for a portion of entries. This dataset is generated using a curated pipeline leveraging Curator.LLM() and Curator.CodeExecutor(). Each entry consists of a mathematical question, a detailed script for an educational animation, and the Manim code required to render the animation.

Data Generation Process

The dataset was generated using the following structured pipeline:

1. Subject & Topic Selection: A diverse set of subjects was chosen, covering a range of difficulty levels. Various topics were identified within each subject.
2. Question Generation: For each topic, multiple unique questions were generated, resulting in a total of 1,000 questions.
3. Script Creation: Each question was paired with a detailed script designed to guide the generation of a math animation video.
4. Manim Code Generation: Using Claude-Sonnet-3.7 with thinking, Manim code snippets were created for each question-script pair.

Rendering the Animations: The generated Manim code was executed using Curator’s CodeExecutor with Docker backend and the Manim Docker image.

<img src="https://cdn.prod.website-files.com/674dabf3c2544460320f97c3/67bf4ff8b797cab695dddef6_Bespoke%20Manim%20Pipeline%20code%20exectuor.png" width="600"/>

To get started, please check the docs and the following examples:

1. [Math video generation](https://github.com/bespokelabsai/curator/tree/main/examples/code-execution/math-animation)
2. [Colab example](https://colab.research.google.com/drive/1YKj1-BC66-3LgNkf1m5AEPswIYtpOU-k?usp=sharing)

For feature requests and bugs, please raise an issue in [Curator repository](https://github.com/bespokelabsai/curator/issues) in GitHub.

<a href="https://github.com/bespokelabsai/curator/"> <img src="https://huggingface.co/datasets/bespokelabs/Bespoke-Stratos-17k/resolve/main/made_with_curator.png" alt="基于Curator构建" width=200px> </a>

Bespoke-Manim 是一个完全合成的数据集，包含1000个数学动画视频脚本，其中部分条目附带由Manim生成的对应动画。本数据集依托Curator.LLM()与Curator.CodeExecutor()构建的精选流水线生成。每条数据均包含一道数学问题、一份用于教学动画的详细脚本，以及渲染该动画所需的Manim代码。

数据生成流程

1. 主题与选题选取：选取覆盖不同难度层级的多元主题域，并在每个主题域下细分出具体研究话题。
2. 问题生成：针对每个细分话题生成多个独特问题，最终累计得到1000道问题。
3. 脚本创作：为每道问题匹配一份详细脚本，用于指导数学动画视频的生成。
4. Manim代码生成：借助支持思维链的Claude-Sonnet-3.7，为每道问题-脚本对生成Manim代码片段。

动画渲染流程：生成的Manim代码通过Curator的CodeExecutor执行，该工具依托Docker后端与Manim Docker镜像运行。

<img src="https://cdn.prod.website-files.com/674dabf3c2544460320f97c3/67bf4ff8b797cab695dddef6_Bespoke%20Manim%20Pipeline%20code%20executor.png" alt="Bespoke-Manim 流水线代码执行流程示意图" width=600/>

若需快速上手，请查阅官方文档及以下示例：
1. [数学动画生成示例](https://github.com/bespokelabsai/curator/tree/main/examples/code-execution/math-animation)
2. [Colab 实操示例](https://colab.research.google.com/drive/1YKj1-BC66-3LgNkf1m5AEPswIYtpOU-k?usp=sharing)

若需提交功能需求或反馈bug，请前往GitHub上的[Curator代码仓库](https://github.com/bespokelabsai/curator/issues)发起议题。

搜集汇总

背景与挑战

背景概述

Bespoke-Manim是一个包含1000个数学动画视频脚本的合成数据集，部分条目配有Manim生成的动画。该数据集通过Curator.LLM()和Curator.CodeExecutor()的流程生成，涵盖多种数学主题和难度级别，适用于数学视频生成和教育用途。