manojdahal191gom/claude-opus-4.6-4.7-reasoning-8.7k

--- license: apache-2.0 task_categories: - text-generation - question-answering language: - en tags: - sft - chain-of-thought - coding - math - roleplay - science - humanities - art - multi-turn - text - json pretty_name: Claude Opus 4.6/4.7 Reasoning Dataset size_categories: - 1K<n<10K --- # Background Ended up with some tokens to burn on a Claude Max plan. Assembly began during 4.6 and moved to 4.7. Model is tagged. The development evolved as it went along. The dataset has not been manually reviewed. It's entirely Claude developed. # Clarification on Reasoning The reasoning is not Claude's actual chain-of-thought (cot) and is not summarized cot. It's a fully synthetic cot created as part of the Assistant response to mimic the type of "thinking" expected to reach the Assistant response. There are now non-reasoning versions as well. ## Files Four datasets provided: | Split | File | Examples | Contents | |-------|------|---------:|----------| | **Full** | `full_train.jsonl` | 8,706 | All examples across all 28 categories. | | **Instruct** | `instruct_train.jsonl` | 7,217 | All 24 instructional categories — coding, math, sciences, humanities, arts, finance, medicine, law, business, linguistics, creative writing, general. | | **Roleplay** | `roleplay_train.jsonl` | 1,489 | The four creative categories — `roleplay_hero`, `roleplay_villain`, `roleplay_crossover`, `narrative_prose`. | | **Code** | `code_train.jsonl` | 1,840 | `coding` + `math` only. For coding/math-focused fine-tunes. | # SLOP Readme A synthetic instruction-tuning dataset designed to teach language models **how to think**, not just what to say. Every example is written to expert standards across 28 categories spanning coding, math, the sciences, the humanities, the arts, finance, medicine, law, business, linguistics, creative writing, roleplay, and narrative prose. Every assistant turn includes a `<think>` block — genuine deliberation, not a reformatted answer. ## Dataset Summary | Metric | Value | |--------|-------| | **Total examples** | 8,706 | | **Estimated tokens** | ~17.0M | | **Categories** | 28 (all populated) | | **With reasoning tags** | 8,706 (100%) | | **Multi-turn conversations** | 3,454 (39.7%) | | **Unique system prompts** | 5,814 | | **Format** | OpenAI chat (JSONL) | | **Language** | English | | **Teacher models** | Claude Opus 4.6 (53.7%), Claude Opus 4.7 (46.3%) | ## What Makes This Dataset Different - **Genuine chain-of-thought reasoning, on every example.** Each assistant turn opens with a `<think>...</think>` block of 150–500 words showing the model considering multiple angles, weighing alternatives, and planning response structure — not "Step 1, Step 2" reformulations of the answer. - **Expert-level depth.** Responses are written to satisfy senior practitioners. Coding answers explain design trade-offs. History answers engage with historiographical debate. Creative-writing critique includes line-level rewrites. Roleplay characters have internally coherent worldviews. - **Natural user voice.** User messages sound like real people — frustrated developers pasting broken code, students challenging an explanation, novelists stuck mid-draft, editors asking for a tonal shift. Hard rule: at most ~20% of user messages start with *What* or *How*. - **5,814 unique system prompts.** Domain-specific personas (e.g. *"You are a database performance consultant working on a Postgres query that's hitting timeouts under load"*) rather than one generic *"helpful assistant"* repeated thousands of times. - **Character-accurate roleplay.** Roleplay examples are built around source-material voice, verbal habits, and worldview — not surface costumes. Includes a deliberately *dark* track of villain, hero, and crossover examples written in the literary register of Le Carré, McCarthy, Atwood, Tartt, Flynn, Bakker, and similar reference points. - **No refusals or safety hedging.** Refusals, content warnings, and clarification-only turns are intentionally excluded. This dataset is for teaching *capability*, not for replacing alignment training. ## Categories 28 categories grouped into instructional and creative/roleplay sets. All 28 are populated; the largest categories carry the foundational legacy content, while the newer per-discipline categories give per-domain coverage. ### Instructional categories (24) | Category | Description | |----------|-------------| | `coding` | Working code with design trade-offs, debugging, architecture. Python, TypeScript, Go, Rust, SQL, more. | | `math` | Pure and applied mathematics, statistics, probability, geometry, algebra, calculus, logic. | | `physics` | Mechanics, thermodynamics, quantum, relativity, electromagnetism, optics. | | `biology` | Genetics, evolution, ecology, microbiology, neuroscience, cell biology. | | `chemistry` | Organic, inorganic, biochemistry, materials science. | | `earth_science` | Geology, climate, meteorology, oceanography, astronomy, paleontology. | | `science` | General-science catch-all for cross-disciplinary topics. | | `history` | Events, historiography, primary sources, ancient through modern. | | `philosophy` | Epistemology, ethics, logic, metaphysics, aesthetics. | | `psychology` | Cognition, behavior, development, social psychology. | | `political_science` | Governance, international relations, policy, political theory. | | `sociology` | Social structures, institutions, inequality, demography. | | `economics` | Macro/microeconomics, econometrics, development economics, game theory. | | `geography` | Human and physical geography, cartography, geopolitics, urban planning. | | `literature` | Literary criticism, poetry analysis, comparative literature, theory. | | `humanities` | Catch-all for cross-disciplinary humanities topics. | | `arts` | Music, film, theater, painting, sculpture, architecture, photography, design. | | `finance` | Investing, accounting, banking, markets, personal finance, trading. | | `medicine` | Clinical reasoning, pharmacology, anatomy, public health, epidemiology. | | `law` | Constitutional, contracts, criminal, civil, jurisprudence, regulation. | | `business` | Management, strategy, leadership, operations, marketing, entrepreneurship. | | `linguistics` | Translation, etymology, phonetics, grammar, syntax, language acquisition. | | `creative_writing` | Craft-focused coaching with concrete techniques, before/after rewrites, line-level analysis. | | `general` | Practical advice, explanations, life questions. Depth matched to question complexity. | ### Creative / roleplay categories (4) | Category | Description | |----------|-------------| | `roleplay_hero` | Heroic and morally complex protagonists with rich, source-accurate voices. | | `roleplay_villain` | Antagonists with internally coherent worldviews — not cartoonish evil. | | `roleplay_crossover` | Cross-canon character pairings with distinct voices and dramatic dynamics. | | `narrative_prose` | Publishable-quality literary fiction in named author voices (Hemingway, Tolstoy, Austen, Pynchon, McCarthy, Le Carré, etc.) and genres. | ## Overall | Metric | Value | |---|---:| | Examples | 8,706 | | Tokens (estimated) | 17,013,533 | | Avg tokens / example | 1,954 | | With reasoning | 8,706 (100.0%) | | Multi-turn | 3,454 (39.7%) | | Single-turn | 5,252 (60.3%) | ## Category Counts | Category | Examples | Tokens | Multi-turn % | |----------|---------:|-------:|-------------:| | coding | 1,628 | 2,545,221 | 30.4% | | humanities | 862 | 1,849,708 | 32.5% | | science | 737 | 1,681,346 | 37.4% | | roleplay_hero | 419 | 640,084 | 63.5% | | roleplay_villain | 378 | 635,984 | 60.8% | | narrative_prose | 377 | 710,807 | 43.0% | | roleplay_crossover | 315 | 581,188 | 56.8% | | creative_writing | 281 | 532,504 | 30.6% | | medicine | 280 | 519,662 | 22.1% | | biology | 277 | 541,013 | 21.3% | | general | 276 | 284,696 | 37.0% | | arts | 245 | 576,170 | 41.2% | | chemistry | 221 | 508,546 | 52.9% | | physics | 220 | 512,196 | 56.8% | | math | 212 | 394,907 | 54.2% | | geography | 155 | 358,321 | 42.6% | | history | 155 | 348,822 | 41.3% | | economics | 155 | 380,372 | 42.6% | | political_science | 154 | 374,901 | 38.3% | | sociology | 154 | 378,261 | 42.2% | | business | 152 | 315,065 | 38.2% | | earth_science | 152 | 358,209 | 41.4% | | finance | 151 | 328,607 | 38.4% | | philosophy | 150 | 335,514 | 41.3% | | linguistics | 150 | 306,889 | 39.3% | | literature | 150 | 299,606 | 38.7% | | psychology | 150 | 339,565 | 39.3% | | law | 150 | 375,360 | 41.3% | Per-category JSONL splits live in `categories/`. ## By Model Every example carries a `model` field identifying which Claude model generated it. | Model | Count | Share | Tokens | |---|---:|---:|---:| | claude-opus-4-6 | 4,675 | 53.7% | 6,304,169 | | claude-opus-4-7 | 4,031 | 46.3% | 10,709,363 | The two model populations are roughly balanced by example count, but Opus 4.7 examples carry ~70% more tokens on average — newer waves trend toward longer multi-turn content. ## Turn Distribution | Turns | Examples | % | |------:|---------:|---:| | 1 | 5,252 | 60.3% | | 2 | 1,491 | 17.1% | | 3 | 1,858 | 21.3% | | 4 | 82 | 0.9% | | 5 | 21 | 0.2% | | 6 | 2 | 0.0% | Multi-turn conversations are designed to teach models to build on context, handle follow-ups that change direction, defend a craft choice, revise on request, and adjust depth based on user response. ## Response Length Distribution Assistant message length, in characters: | Percentile | Characters | |------------|-----------:| | p10 | 2,061 | | p25 | 2,914 | | Median | 4,239 | | p75 | 5,682 | | p90 | 7,052 | | Max | 30,026 | Reasoning blocks themselves are typically 150–500 words; the rest is the user-facing answer. ## Format Standard OpenAI chat format in JSONL. Each line is one JSON object with `category`, `messages`, and `model` fields: ```json { "category": "coding", "model": "claude-opus-4-7", "messages": [ {"role": "system", "content": "You are a senior backend engineer reviewing performance issues..."}, {"role": "user", "content": "Explain move semantics to me..."}, {"role": "assistant", "content": "<reasoning>\nThe user understands C++ fundamentals but...\n</reasoning>\n\n`std::move` does not move anything. It is a cast..."} ] } ``` The `category` and `model` fields are metadata for filtering and provenance — fine-tuning APIs read only `messages`. ## Terms Use it for things you should use it for but don't use it for anything you shouldn't use it for. Like Anthropic does, always respect terms of use...A synthetic instruction-tuning dataset designed to teach language models how to think, not just what to say. Every example is written to expert standards across 28 categories spanning coding, math, the sciences, the humanities, the arts, finance, medicine, law, business, linguistics, creative writing, roleplay, and narrative prose. Every assistant turn includes a <reasoning> block — genuine deliberation, not a reformatted answer. The dataset features expert-level depth, natural user voice, 5,814 unique system prompts, character-accurate roleplay, and no refusals or safety hedging. It is in OpenAI chat format (JSONL), with 8,706 examples, ~17.0M tokens, 39.7% multi-turn conversations, and generated by Claude Opus 4.6 (53.7%) and 4.7 (46.3%) models.

搜集汇总

数据集介绍

构建方式

该数据集由Claude Opus 4.6与4.7模型生成，数据集构建始于4.6版本，并延续至4.7版本。所有示例均由Claude自主开发，未经人工校审。数据集采用OpenAI聊天格式，以JSONL文件存储，每条记录包含类别、消息序列与模型标识三个字段。消息序列遵循系统提示、用户消息与助手回复的结构，其中助手回复内嵌<think>推理标签。数据集共包含8,706个示例，覆盖28个类别，并依据内容性质划分为完整集、指令集、角色扮演集与代码集四个子集，便于按需选用。

特点

该数据集的核心特色在于每个助手回复均包含一段150至500词的<think>推理区块，展示了模型多角度思考、权衡备选方案及规划回复结构的真实推演过程，而非简单的步骤罗列。数据集涵盖编码、数学、自然科学、人文、艺术、金融、法律等28个专业类别，回复达到专家级深度。用户消息自然真实，避免了“What”或“How”开头的机械句式。拥有5,814个独特系统提示，赋予模型领域特定人格。角色扮演示例忠实于原始材料，并刻意排除了拒绝回答与安全警告，专注于能力培养。

使用方法

用户可通过加载JSONL文件直接使用该数据集进行监督式微调。每条记录中的消息字段可直接被微调API读取，类别与模型字段可用于数据筛选与溯源。数据集提供了四种预划分文件：完整训练集(8,706条)、指令集(7,217条)、角色扮演集(1,489条)与代码集(1,840条)。此外，每个类别还提供了独立的JSONL子文件。用户可根据训练目标选择相应子集，例如针对编码与数学能力提升可选用代码集，或根据模型来源筛选特定Claude版本生成的数据。

背景与挑战

背景概述

该数据集名为Claude Opus 4.6/4.7 Reasoning Dataset，由研究人员利用Claude Max计划中的剩余配额构建，自Claude 4.6版本始，延伸至4.7版本。核心研究问题聚焦于如何通过合成推理数据提升语言模型的深度思考能力，而非仅关注表面输出。数据集涵盖编码、数学、科学、人文、艺术等28个类别，包含约8,700条样本，总计约17M tokens。其独特之处在于，每条助手回复均嵌入一个真正的链式推理（<think>）模块，模拟人类多角度权衡与结构规划过程。该数据集对指令微调领域具有重要影响力，尤其为培养模型的专家级深度分析能力提供了高质量训练材料。

当前挑战

数据集面临的挑战包括三个方面。首先，在领域问题层面，传统指令微调数据集往往侧重于“说什么”，而忽视了“如何思考”，导致模型缺乏真正的推理能力。该数据集通过合成推理模块解决了这一局限，但如何确保合成的推理过程真实反映人类认知仍是一大难题。其次，在构建过程中，数据集由单一模型Claude Opus自动生成，未经人工审查，可能引入模型自身的偏差或错误。同时，数据集中避开了拒绝回答和安全缓冲等涉及安全训练的样本，这虽然专注于提升能力，但可能影响模型在实际应用中的安全鲁棒性。最后，各类别样本数量分布不均（如编码类1,628条，而多数类别仅约150条），可能造成微调过程中的领域偏向，挑战平衡性设计。

常用场景

经典使用场景

该数据集是面向语言模型深层推理能力构建的合成指令微调数据集，涵盖编程、数学、自然科学、人文学科、医学、法律、金融、创意写作、角色扮演等28个专业领域。其核心使用场景在于训练大语言模型习得‘如何思考’的能力，而非简单地记忆输出模式。每条助手回复的开端均嵌入一段150至500词的<think>思维链，模拟模型在回答问题前的多角度权衡、逻辑推演与结构规划过程。这种设计使模型能够将推理过程与最终回答有机分离，从而在对话中展现出更加审慎、具有深度且符合专业标准的回应质量。

解决学术问题

在学术研究层面，该数据集旨在破解当前指令微调范式下语言模型‘知其然而不知其所以然’的困境。传统微调往往强调最终答案的准确性，却忽视了中间推理过程对模型泛化能力与鲁棒性的决定性作用。通过提供全量、真实模拟的思维链数据，该数据集有效支撑了关于推理蒸馏、思维链迁移学习以及多步推理可解释性的前沿探索。研究者在分析模型如何在不同学科场景中构建推理路径时，能够借此填补合成思维链与人类认知过程之间的映射鸿沟，推动推理增强型语言模型的理论建构。

衍生相关工作

该数据集衍生的相关工作集中体现在推理增强型语言模型的设计与评估方法上。基于其全量思维链标注的特性，研究者可以开发出面向推理质量的自动评价指标，如推理逻辑的一贯性、假设空间覆盖度及反驳性自洽性等。同时，该数据集为多轮交互中推理能力的持续优化提供了训练素材，催生了诸如上下文感知推理规划、分步推理回退机制以及思维链剪枝与增强等算法研究。此外，不同模型版本（Claude Opus 4.6与4.7）的混合数据也支持跨代模型的推理能力对比与归因分析，为新一代推理建模技术的发展奠定了数据基础。