yesaf990/AIME-God-v5-Pure-Reasoning
收藏Hugging Face2026-03-18 更新2026-03-29 收录
下载链接:
https://hf-mirror.com/datasets/yesaf990/AIME-God-v5-Pure-Reasoning
下载链接
链接失效反馈官方服务:
资源简介:
# 🔱 AIME-God v5.2 "Zero-System" (Training)
The ultimate dataset for fine-tuning models to master the AIME/Olympiad math competitions, specifically targeting the hardest **Tier-15** (problems 11-15).
## 🚀 Key Features
- **Zero-System Strategy**: System instructions are merged into the first User turn for 100% compatibility with Vertex AI Induction and Managed SFT.
- **Tier-15 Mastery**: Includes high-density examples for LTE Lemma, Geometric Inversion, Barycentric Coordinates, Generating Functions, and Roots of Unity Filters.
- **Reasoning Depth**: 50,000 rows of pure, high-signal reasoning blocks with explicit `<think>` steps.
## 📊 Dataset Structure
- **Format**: Conversational JSONL (`messages` list)
- **Roles**: `user`, `assistant` (System prompt merged into first user message).
- **Target**: Qwen-3-8B / Llama-3-8B fine-tuning.
## ⚙️ Recommended Training Config
- **Epochs**: 2
- **Adapter**: LoRA (Rank 16)
- **Learning Rate**: 1e-5
# 🔱 AIME-God v5.2 "零样本系统"(训练数据集)
本数据集为面向AIME与奥林匹克数学竞赛的模型微调专用终极数据集,专门针对难度最高的**Tier-15**(第11至15题)题型设计。
## 🚀 核心特性
- **零样本系统策略**:将系统提示词合并至首个用户对话轮次,100%兼容Vertex AI Induction与托管式监督微调(Managed SFT)。
- **Tier-15题型精通**:涵盖大量针对LTE引理(LTE Lemma)、几何反演(Geometric Inversion)、重心坐标系(Barycentric Coordinates)、生成函数(Generating Functions)以及单位根滤波器(Roots of Unity Filters)的高密度训练示例。
- **推理深度保障**:包含50000条高纯度、高有效信号的推理块,带有显式`<think>`思考步骤。
## 📊 数据集结构
- **格式**:对话式JSONL格式(包含`messages`列表字段)
- **角色定义**:`user`(用户)、`assistant`(助手)(系统提示词已合并至首个用户消息内)
- **适配模型**:Qwen-3-8B / Llama-3-8B 微调任务
## ⚙️ 推荐训练配置
- **训练轮次**:2轮
- **适配器**:LoRA(秩为16)
- **学习率**:1×10^-5
提供机构:
yesaf990



