vanty120/Gpt-5.4-Xhigh-Reasoning-2000x
收藏资源简介：
--- language: - en license: apache-2.0 task_categories: - question-answering - text-generation size_categories: - 1K<n<10K tags: - reasoning - math - code - science - distillation - chain-of-thought - gpt-5.4 - gemini-3.1-pro - thinking - sft - hard-reasoning pretty_name: Gpt-5.4-Xhigh-Reasoning-2750x --- # Gpt-5.4-Xhigh-Reasoning-2750x A premium-quality reasoning dataset containing **2,752 elite samples** distilled from **GPT-5.4 XHIGH** (the highest reasoning effort tier of GPT-5.4). Each sample features deep, multi-step Chain-of-Thought traces that are significantly longer and more rigorous than standard GPT-5.4 outputs. This dataset is specifically designed for **Supervised Fine-Tuning (SFT)** to transform general-purpose language models into powerful reasoning models with explicit thinking capabilities. ## Dataset Summary | Property | Value | |---|---| | **Total Samples** | 2,752 | | **Teacher Model** | GPT-5.4 XHIGH (Maximum Reasoning Effort) | | **Seed Data** | [Qwen3.5-reasoning-700x](https://huggingface.co/datasets/Jackrong/Qwen3.5-reasoning-700x) + [Opus-4.6-Reasoning-3000x-filtered](https://huggingface.co/datasets/nohurry/Opus-4.6-Reasoning-3000x-filtered) + [gemini-3.1-pro-hard-high-reasoning](https://huggingface.co/datasets/Roman1111111/gemini-3.1-pro-hard-high-reasoning) | | **Language** | English | | **Domains** | Mathematics, Code, Science, STEM, Security, Economics, and 60+ expert-level domains | | **Avg. Thinking Length** | ~12,600 characters per sample | ### Why XHIGH? GPT-5.4 supports multiple reasoning effort levels. **XHIGH** is the maximum tier, which forces the model to allocate significantly more compute to its internal chain-of-thought before producing a final answer. This results in: - **Deeper logical decomposition** compared to standard GPT-5.4 outputs - **More self-correction steps** within the reasoning trace - **Higher accuracy** on complex multi-step problems ## Seed Data Sources ### Source 1: Standard Reasoning (2,007 samples) High-quality prompts sourced from [Qwen3.5-reasoning-700x](https://huggingface.co/datasets/Jackrong/Qwen3.5-reasoning-700x) and [Opus-4.6-Reasoning-3000x-filtered](https://huggingface.co/datasets/nohurry/Opus-4.6-Reasoning-3000x-filtered), covering math, code, science, and instruction-following tasks. ### Source 2: Gemini 3.1 Pro Hard Reasoning (745 samples) Ultra-hard prompts sourced from [Roman1111111/gemini-3.1-pro-hard-high-reasoning](https://huggingface.co/datasets/Roman1111111/gemini-3.1-pro-hard-high-reasoning). These prompts were originally generated by an agentic workflow (Gemini 3 Flash as orchestrator) using the following high-intensity system instruction: > **System Instruction**: Act as a "Super-Intelligence Evaluator". Generate distinct, complex but solvable prompts that require extreme logic. > > **Requirements**: > 1. **Difficulty**: The question must be unsolvable by simple retrieval. It requires multi-step logic, derivation, or synthesis of conflicting information. > 2. **Concept**: Pick a specific, niche concept within the target domain. > 3. **Prompt Text**: The user prompt should be detailed (code snippets, math proofs, or philosophical paradoxes). > 4. **No Fluff**: Go straight to the hard part. This source spans 60+ expert-level domains including: - **Physics**: QFT, General Relativity, Condensed Matter, Thermodynamics - **Math**: Algebraic Topology, Analytic Number Theory, Category Theory - **Coding/CS**: CUDA/OpenCL HPC, Database Internals, LLVM IR, ZK-Proofs - **Biology/Med**: CRISPR Off-target Analysis, Protein Folding, Pharmacokinetics - **Security**: Prompt Injection Defense, Cryptanalysis, Side-channel Attacks - **Strategic Logic**: Game Theory, Supply Chain Crisis Modeling, Urban Planning - **Benchmarks**: ARC-AGI, LiveCodeBench v6, TheoremQA, MathVista ## Domain Distribution | Category | Count | Percentage | |---|---|---| | Mathematics | 1,581 | 57.4% | | Code | 174 | 6.3% | | Science | 136 | 4.9% | | Instruction Following | 116 | 4.2% | | Prompt Injection & Jailbreak Defense | 70 | 2.5% | | Algebraic Topology | 65 | 2.4% | | Bioinformatics Algorithms | 43 | 1.6% | | Computational Chemistry (DFT) | 41 | 1.5% | | Other Expert Domains (60+) | 526 | 19.1% | ## Difficulty Distribution | Difficulty | Count | Description | |---|---|---| | Medium | 1,986 | Undergraduate level | | Hard | 188 | Professional / competition level | | Extreme | 111 | Research frontier | | Expert | 124 | PhD-level, research-grade problems | | Advanced+ | 343 | Advanced, Super-Intelligence, Graduate, Olympiad | ## Dataset Structure Each sample contains the following fields: ```json { "category": "Algebraic Topology", "difficulty": "Extreme", "instruction": "The original question or problem statement...", "thinking": "Full chain-of-thought reasoning trace from GPT-5.4 XHIGH...", "response": "The final, polished answer..." } ``` | Field | Description | |---|---| | `category` | Domain classification (60+ categories) | | `difficulty` | Difficulty tier: `medium`, `hard`, `extreme`, `expert`, `advanced`, etc. | | `instruction` | The original problem or question | | `thinking` | Complete reasoning trace (Chain-of-Thought) from GPT-5.4 XHIGH | | `response` | Final solution / answer | ## Generation Pipeline 1. **Seed Selection**: High-quality prompts sourced from three complementary datasets covering standard reasoning (math, code, science) and ultra-hard expert-level domains (60+ fields). 2. **Distillation**: Each prompt was processed through **GPT-5.4** with `reasoning_effort=xhigh`, extracting both the internal reasoning trace and the final output. 3. **Quality Control**: Samples with empty thinking or responses were filtered out. Prompt injection artifacts were cleaned from the input. ### Training Format (ChatML with Thinking) ``` <|im_start|>system You are a helpful assistant that thinks step-by-step.<|im_end|> <|im_start|>user {instruction}<|im_end|> <|im_start|>assistant <thinking> {thinking} </thinking> {response}<|im_end|> ``` ## Disclaimers - **LLM Hallucinations**: While GPT-5.4 XHIGH produces highly rigorous outputs, a small number of reasoning errors may still exist. Sample inspection before fine-tuning is recommended. - **Expert Verification**: The hard-reasoning subset contains solutions so technical that they may require subject-matter experts (PhDs) to verify accuracy. - **License**: This dataset is released under the Apache 2.0 license. Usage must comply with [OpenAI's Terms of Service](https://openai.com/policies/terms-of-use). ## Credits - **Teacher Model**: [GPT-5.4](https://openai.com/gpt-5) by OpenAI (XHIGH reasoning effort) - **Seed Datasets**: - [Jackrong/Qwen3.5-reasoning-700x](https://huggingface.co/datasets/Jackrong/Qwen3.5-reasoning-700x) (Alibaba-Superior-Reasoning-Stage2) - [nohurry/Opus-4.6-Reasoning-3000x-filtered](https://huggingface.co/datasets/nohurry/Opus-4.6-Reasoning-3000x-filtered) - [Roman1111111/gemini-3.1-pro-hard-high-reasoning](https://huggingface.co/datasets/Roman1111111/gemini-3.1-pro-hard-high-reasoning) - **Distillation Pipeline**: Built by [vanty120](https://huggingface.co/vanty120)
--- 语言：英语 许可证：Apache 2.0 任务类别： - 问答 - 文本生成 样本规模：1000 < 样本量 < 10000 标签： - 推理 - 数学 - 代码 - 科学 - 知识蒸馏 - 思维链（Chain-of-Thought） - GPT-5.4 - Gemini-3.1-Pro - 思考 - 监督微调（SFT） - 硬核推理 美观名称：GPT-5.4-Xhigh-Reasoning-2750x --- # GPT-5.4-Xhigh-Reasoning-2750x 本数据集为高质量推理数据集，包含2752条精选样本，均从GPT-5.4 XHIGH（GPT-5.4的最高推理算力档位）中蒸馏得到。每条样本均包含深度多步思维链（Chain-of-Thought）推理轨迹，相较于标准GPT-5.4输出，其长度更长、严谨性更强。 本数据集专为监督微调（SFT）设计，可用于将通用大语言模型（Large Language Model，LLM）改造为具备显式思考能力的高性能推理模型。 ## 数据集概览 | 数据集属性 | 取值 | |---|---| | **总样本量** | 2752 | | **教师模型** | GPT-5.4 XHIGH（最高推理算力档位） | | **种子数据集** | [Qwen3.5-reasoning-700x](https://huggingface.co/datasets/Jackrong/Qwen3.5-reasoning-700x) + [Opus-4.6-Reasoning-3000x-filtered](https://huggingface.co/datasets/nohurry/Opus-4.6-Reasoning-3000x-filtered) + [gemini-3.1-pro-hard-high-reasoning](https://huggingface.co/datasets/Roman1111111/gemini-3.1-pro-hard-high-reasoning) | | **语言** | 英语 | | **覆盖领域** | 数学、代码、科学、STEM、安全、经济学，以及60余个专家级细分领域 | | **平均推理轨迹长度** | 单样本约12600字符 | ### 为何选用XHIGH档位？ GPT-5.4提供多种推理算力档位，**XHIGH**为最高档位，可迫使模型在生成最终答案前，为内部思维链推理分配更多计算资源。这一设计可带来如下优势： - 相较于标准GPT-5.4输出，具备更深入的逻辑拆解能力 - 推理轨迹中包含更多自我修正步骤 - 在复杂多步问题上具备更高的求解准确率 ## 种子数据集来源 ### 来源1：标准推理样本（2007条） 本部分样本的高质量提示词源自[Qwen3.5-reasoning-700x](https://huggingface.co/datasets/Jackrong/Qwen3.5-reasoning-700x)与[Opus-4.6-Reasoning-3000x-filtered](https://huggingface.co/datasets/nohurry/Opus-4.6-Reasoning-3000x-filtered)，覆盖数学、代码、科学以及指令遵循类任务。 ### 来源2：Gemini 3.1 Pro 硬核推理样本（745条） 本部分样本的超硬核提示词源自[Roman1111111/gemini-3.1-pro-hard-high-reasoning](https://huggingface.co/datasets/Roman1111111/gemini-3.1-pro-hard-high-reasoning)。这些提示词最初由基于智能体的工作流（以Gemini 3 Flash作为调度器）生成，采用如下高强度系统指令： > **系统指令**：扮演“超级智能评估者”，生成独特、复杂但可解的提示词，且需具备极强的逻辑推理要求。 > > **生成要求**： > 1. **难度要求**：问题无法通过简单检索解决，需借助多步逻辑推理、推导或冲突信息整合完成解答 > 2. **概念要求**：选取目标领域内特定的细分概念 > 3. **提示词文本**：用户提示词需具备足够细节（可包含代码片段、数学证明或哲学悖论） > 4. **无冗余内容**：直接切入核心难点 本来源覆盖60余个专家级细分领域，具体包括： - **物理学**：量子场论（QFT）、广义相对论、凝聚态物理、热力学 - **数学**：代数拓扑、解析数论、范畴论 - **编程/计算机科学**：CUDA/OpenCL 高性能计算、数据库内核、LLVM 中间表示、零知识证明（ZK-Proofs） - **生物学/医学**：CRISPR 脱靶分析、蛋白质折叠、药代动力学 - **安全领域**：提示词注入防御、密码分析、侧信道攻击 - **战略逻辑**：博弈论、供应链危机建模、城市规划 - **基准测试集**：ARC-AGI、LiveCodeBench v6、TheoremQA、MathVista ## 领域分布 | 领域类别 | 样本数量 | 占比 | |---|---|---| | 数学 | 1581 | 57.4% | | 代码 | 174 | 6.3% | | 科学 | 136 | 4.9% | | 指令遵循 | 116 | 4.2% | | 提示词注入与越狱防御 | 70 | 2.5% | | 代数拓扑 | 65 | 2.4% | | 生物信息学算法 | 43 | 1.6% | | 计算化学（密度泛函理论，DFT） | 41 | 1.5% | | 其他专家级领域（共60+个） | 526 | 19.1% | ## 难度分布 | 难度档位 | 样本数量 | 难度说明 | |---|---|---| | 中等 | 1986 | 本科阶段难度 | | 困难 | 188 | 专业/竞赛级难度 | | 极难 | 111 | 研究前沿级难度 | | 专家级 | 124 | 博士级、研究级问题 | | 进阶+ | 343 | 高阶、超级智能级、研究生/奥赛级难度 | ## 数据集结构 每条样本包含如下字段： json { "category": "代数拓扑", "difficulty": "极难", "instruction": "原始问题或任务描述...", "thinking": "GPT-5.4 XHIGH生成的完整思维链推理轨迹...", "response": "最终打磨完成的答案..." } | 字段名 | 字段说明 | |---|---| | `category` | 领域分类（共60+个类别） | | `difficulty` | 难度档位：`中等`、`困难`、`极难`、`专家级`、`进阶`等 | | `instruction` | 原始问题或任务描述 | | `thinking` | GPT-5.4 XHIGH生成的完整推理轨迹（思维链） | | `response` | 最终解决方案/答案 | ## 数据生成流程 1. **种子数据集选取**：从三个互补的数据集获取高质量提示词，覆盖标准推理任务（数学、代码、科学）以及60+个领域的超硬核专家级任务。 2. **知识蒸馏**：将每条提示词输入至设置为`reasoning_effort=xhigh`的**GPT-5.4**模型，提取其内部推理轨迹与最终输出结果。 3. **质量控制**：过滤掉推理轨迹或答案为空的样本，清理输入中的提示词注入残留内容。 ### 训练格式（带思维链的ChatML格式） <|im_start|>system 你是一位会逐步思考的实用助手。<|im_end|> <|im_start|>user {instruction}<|im_end|> <|im_start|>assistant <thinking> {thinking} </thinking> {response}<|im_end|> ## 免责声明 - **大语言模型幻觉问题**：尽管GPT-5.4 XHIGH的输出具备较高严谨性，但仍可能存在少量推理错误。建议在微调前对样本进行人工检查。 - **专家验证需求**：硬核推理子集的解决方案技术门槛较高，可能需要领域专家（如博士研究者）才能验证其准确性。 - **许可证**：本数据集采用Apache 2.0许可证发布。使用时需遵守[OpenAI服务条款](https://openai.com/policies/terms-of-use)。 ## 致谢 - **教师模型**：OpenAI出品的[GPT-5.4](https://openai.com/gpt-5)（采用XHIGH推理档位） - **种子数据集**： - [Jackrong/Qwen3.5-reasoning-700x](https://huggingface.co/datasets/Jackrong/Qwen3.5-reasoning-700x)（Alibaba-Superior-Reasoning-Stage2） - [nohurry/Opus-4.6-Reasoning-3000x-filtered](https://huggingface.co/datasets/nohurry/Opus-4.6-Reasoning-3000x-filtered) - [Roman1111111/gemini-3.1-pro-hard-high-reasoning](https://huggingface.co/datasets/Roman1111111/gemini-3.1-pro-hard-high-reasoning) - **知识蒸馏流水线**：由[vanty120](https://huggingface.co/vanty120)搭建
