Annoy-PyEdu-Rs
收藏资源简介:
Annoy-PythonEdu-Rs 是一个专注于 Python 教育场景的数据集,属于完整数据集的一个子集。该数据集通过完全基于大语言模型的方法构建,使用 DeepSeek-V2.5 合成所有响应,以克服传统方法中难以获得确定性反向函数以及自动构建轨迹受限于预设计模板、缺乏自由形式自然语言推理表达力的问题。数据集可能包含 Python 代码片段及其对应的执行轨迹或自然语言推理过程,旨在支持代码执行预测、程序理解或教育场景下的推理任务。由于合作者合规要求,目前仅公开此 PythonEdu 子集。
Annoy-PythonEdu-Rs is a dataset focused on Python education scenarios, a subset of the complete dataset. It is constructed entirely through a large language model approach, using DeepSeek-V2.5 to synthesize all responses, to overcome the difficulty of obtaining deterministic inverse functions in traditional methods and the limitations of automatically constructed trajectories being constrained by predefined templates and lacking free-form natural language reasoning expressiveness. The dataset may contain Python code snippets along with their corresponding execution trajectories or natural language reasoning processes, aiming to support code execution prediction, program understanding, or reasoning tasks in educational scenarios. Due to collaborator compliance requirements, only this PythonEdu subset is currently publicly available.
数据集概述
基本信息
- 数据集名称:Annoy-PythonEdu-Rs(PythonEdu-Rs 子集)
- 所属项目:Annoy(论文标题)
- 数据来源:基于 DeepSeek-V2.5 模型合成
- 发布说明:因合作方合规要求,仅公开完整数据集的 PythonEdu-Rs 子集
数据构建背景
- 虽然完整的可执行代码理论上可生成可靠的执行轨迹,但存在两个现实挑战:
- 获取用于输入预测的确定性反向函数不切实际
- 自动构建的轨迹受限于预设模板,缺乏自由形式自然语言推理的表达力和泛化能力
- 为此,该项目采用 DeepSeek-V2.5 进行完全基于 LLM 的合成方法来生成所需响应,理由是其性能顶尖且成本远低于其他先进 LLM
相关资源
原始数据
- 处理后的原始数据:yfeng123456/Annoy-PyEdu-Rs-Raw(Hugging Face 链接)
相关模型
该项目采用多种基础模型进行训练,分为两种方法(Annoy 与 Annoy++)和两个训练阶段(Stage 1、Stage 2):
| 基础模型 | Annoy Stage 1 | Annoy Stage 2 | Annoy++ Stage 1 | Annoy++ Stage 2 |
|---|---|---|---|---|
| Qwen 2.5 7B Coder | yfeng123456/qwen2.5-7b-coder_spec_stage1 | yfeng123456/qwen2.5-7b-coder_spec | yfeng123456/qwen2.5-7b-coder_spec_pp_stage1 | yfeng123456/qwen2.5-7b-coder_spec_pp |
| LLaMA 3.1 8B | yfeng123456/llama3.1-8b_spec_stage1 | yfeng123456/llama3.1-8b_spec | yfeng123456/llama3.1-8b_spec_pp_stage1 | yfeng123456/llama3.1-8b_spec_pp |
| DeepSeek v2 Lite Coder | yfeng123456/dsv2-lite-coder_spec_stage1 | yfeng123456/dsv2-lite-coder_spec | yfeng123456/dsv2-lite-coder_spec_pp_stage1 | yfeng123456/dsv2-lite-coder_spec_pp |
引用入口
- 论文链接(Hugging Face Papers)
- 项目主页:specx.github.io
- 发布资源合集:Hugging Face Collections(yfeng123456/specx)
- 代码仓库:github.com/jackmonsignmo/Annoy




