Glint-Research/Opus-4.6-Reasoning-2160x

---
license: apache-2.0
language:
- en
task_categories:
- text-generation
- question-answering
tags:
- reasoning
- chain-of-thought
- distillation
- claude
- opus
- math
- code
- sft
size_categories:
- 1K<n<10K
---

# Opus-4.6-Reasoning-2160x

**2,160 high-quality reasoning traces** generated by Claude Opus 4.6 via OpenRouter, covering mathematics, competitive programming, logic, science, and language tasks. Each example includes the full problem, an extended chain-of-thought, and a final solution — making the dataset suitable for supervised fine-tuning, chain-of-thought distillation, and reasoning-capability transfer to smaller models.

Originally generated as a batch of 3,305 examples; 1,145 were removed during quality filtering (refusals, empty responses, truncated problems). The 2,160 remaining examples are all substantive, complete reasoning demonstrations.

---

## Dataset at a Glance

| Property | Value |
|---|---|
| Rows | 2,160 |
| Avg tokens per row | ~758 |
| Total completion tokens | 1,636,368 |
| Difficulty classes | easy / medium / hard |
| Category classes | 2 (math/reasoning, code/logic) |
| License | Apache 2.0 |
| Generation cost | ~$409 USD at Opus 4.6 pricing |

---

## Schema

```jsonl
{
"id": "drive_minimax_m2.1_questions_70109",
"problem": "Your solution must read input from standard input...",
"thinking": "Looking at what was provided, this appears to be...",
"solution": "The problem statement appears to be incomplete...",
"difficulty": "medium",
"category": "code",
"timestamp": "2026-02-12T21:49:00Z",
"hash": "a3f9c1d2e4b78901"
}
```

| Field | Type | Description |
|---|---|---|
| `id` | string (5–36 chars) | Unique example identifier |
| `problem` | string (5–13.7k chars) | The problem or prompt given to the model |
| `thinking` | string (80–13.8k chars) | Opus 4.6's full chain-of-thought reasoning |
| `solution` | string (20–15.1k chars) | Final answer or solution |
| `difficulty` | string (3 values) | `easy`, `medium`, or `hard` |
| `category` | string (2 values) | High-level domain label |
| `timestamp` | string (32 chars) | ISO 8601 generation timestamp |
| `hash` | string (16 chars) | Deduplication hash |

---

## Loading the Dataset

```python
from datasets import load_dataset

ds = load_dataset("CompactAI-O/Opus-4.6-Reasoning-2160x", split="train")
print(ds[0]["problem"])
print(ds[0]["thinking"])
print(ds[0]["solution"])
```

---

## Training Use Cases

This dataset is designed for **knowledge transfer from a frontier reasoning model to smaller, more efficient models**. Below are the primary training patterns.

### 1. Standard Supervised Fine-Tuning (SFT)

Use `problem` → `solution` for instruction-following training. The solution fields are clean, complete answers suitable for direct target supervision.

```python
def format_sft(row):
return {
"input": row["problem"],
"output": row["solution"]
}
```

Best for: models already capable of basic instruction following that need domain knowledge transfer.

### 2. Chain-of-Thought Distillation

Train on `problem` → `thinking + solution` to transfer Opus 4.6's reasoning style. The `thinking` field contains multi-step reasoning that is not present in shorter/weaker model outputs.

```python
def format_cot(row):
return {
"input": row["problem"],
"output": f"{row['thinking']}\n\n{row['solution']}"
}
```

Best for: models being trained to produce explicit reasoning before answering (Qwen, Phi, Mistral, FANT-class architectures with `<|think|>` tokens).

### 3. Think-Solution Format (FANT3 / structured reasoning)

For architectures that use explicit thinking delimiters (e.g. `<|think|>...<|answer|>...`):

```python
def format_think_answer(row, think_open="<|think|>", think_close="<|/think|>",
ans_open="<|answer|>", ans_close="<|/answer|>"):
return (
f"{think_open}{row['thinking']}{think_close}"
f"{ans_open}{row['solution']}{ans_close}"
)
```

This format is used by FANT3 and similar models that separate latent reasoning from final output at the token level.

### 4. Difficulty-Weighted Sampling

The `difficulty` field allows curriculum learning — train on `easy` first, then progressively introduce `medium` and `hard`.

```python
from datasets import load_dataset

ds = load_dataset("CompactAI-O/Opus-4.6-Reasoning-2160x", split="train")

easy = ds.filter(lambda x: x["difficulty"] == "easy")
medium = ds.filter(lambda x: x["difficulty"] == "medium")
hard = ds.filter(lambda x: x["difficulty"] == "hard")
```

### 5. Category-Filtered Training

Use `category` to mix domain-specific signals with general corpora:

```python
math_ds = ds.filter(lambda x: x["category"] == "math")
code_ds = ds.filter(lambda x: x["category"] == "code")
```

### 6. Sequence-Level KD (Knowledge Distillation)

When used alongside a local teacher model, this dataset supports GKD-style sequence-level distillation. The `thinking` + `solution` sequences serve as teacher samples. No logit alignment is needed — this is pure sequence supervision.

```python
# Example with TRL SFTTrainer
from trl import SFTTrainer, SFTConfig

trainer = SFTTrainer(
model=student_model,
train_dataset=ds,
args=SFTConfig(
output_dir="./output",
max_seq_length=2048,
),
formatting_func=format_cot,
)
trainer.train()
```

---

## Problem Coverage

Examples span a wide range of domains:

- **Mathematics**: algebra, geometry, arithmetic word problems, combinatorics
- **Competitive programming**: string manipulation, graph algorithms, dynamic programming
- **Logic & NLI**: premise/hypothesis entailment, paraphrase detection
- **Science**: chemistry, biology, physics
- **Language tasks**: translation, reading comprehension, extraction
- **General QA**: history, geography, factual recall

---

## Cleaning Report

| Rejection reason | Count | % of original |
|---|---|---|
| Refusal / incomplete problem | 1,090 | 33.0% |
| Empty response | 18 | 0.5% |
| Response had no substance | 17 | 0.5% |
| Problem too short | 14 | 0.4% |
| Response too short | 6 | 0.2% |
| **Total removed** | **1,145** | **34.6%** |
| **Total kept** | **2,160** | **65.4%** |

Cleaning was performed on 2026-02-12. All kept examples have a substantive problem statement and a non-trivial Opus 4.6 response.

---

## Decontamination

Before using this dataset in any evaluation context, apply n-gram decontamination against your eval sets (GSM8K, MATH-500, MMLU). The NuminaMath and similar math-adjacent sources in the original generation prompts carry a small overlap risk. Verbatim contamination rate against standard benchmarks is estimated at <0.5%.

---

## Generation Details

- **Model**: Claude Opus 4.6 via OpenRouter
- **Pricing at generation time**: $5.00/M input tokens, $25.00/M output tokens
- **Total output tokens**: 1,636,368
- **Estimated cost**: ~$409 USD
- **Average turns**: 1.00 (single-turn, no tool calls)
- **Generation date**: February 2026

---

## Citation

If you use this dataset, please credit the original generation effort:

```
@misc{crownelius2026opus46reasoning,
title = {Opus-4.6-Reasoning-2160x},
author = {Crownelius / CompactAI-O},
year = {2026},
url = {https://huggingface.co/datasets/CompactAI-O/Opus-4.6-Reasoning-2160x}
}
```

---

## License

Apache 2.0. Generated outputs from Claude Opus 4.6 are subject to Anthropic's usage policies. Commercial use of distilled model outputs should comply with Anthropic's terms of service.

---
license: apache-2.0
language:
- en
task_categories:
- 文本生成
- 问答任务
tags:
- 推理
- 思维链（Chain-of-Thought）
- 模型蒸馏（Distillation）
- Claude
- Opus
- 数学
- 代码
- 监督微调（Supervised Fine-Tuning，SFT）
size_categories:
- 1K<n<10K
---

# Opus-4.6-Reasoning-2160x

**2160条高质量推理轨迹**由Claude Opus 4.6通过OpenRouter生成，涵盖数学、竞赛编程、逻辑、科学与语言类任务。每条样本均包含完整题目、完整展开的思维链推理过程与最终答案，可适用于监督微调、思维链蒸馏以及将推理能力迁移至小型模型等场景。

该数据集最初生成了3305条样本，经质量过滤后移除了1145条不合格样本（包括模型拒答、空响应、题目截断的样本），最终保留的2160条样本均为内容充实、推理过程完整的演示案例。

---

## 数据集概览

| 属性 | 取值 |
|---|---|
| 样本条数 | 2,160 |
| 单条样本平均Token数 | ~758 |
| 总完成Token数 | 1,636,368 |
| 难度分级 | 简单 / 中等 / 困难 |
| 类别分级 | 2大类（数学/推理、代码/逻辑） |
| 许可证 | Apache 2.0 |
| 生成成本 | 按Opus 4.6定价估算约$409 USD |

---

## 数据集结构

jsonl
{
"id": "drive_minimax_m2.1_questions_70109",
"problem": "Your solution must read input from standard input...",
"thinking": "Looking at what was provided, this appears to be...",
"solution": "The problem statement appears to be incomplete...",
"difficulty": "medium",
"category": "code",
"timestamp": "2026-02-12T21:49:00Z",
"hash": "a3f9c1d2e4b78901"
}

| 字段名 | 数据类型 | 字段说明 |
|---|---|---|
| `id` | 字符串（5–36字符） | 唯一样本标识符 |
| `problem` | 字符串（5–13.7k字符） | 提交给模型的题目或提示词 |
| `thinking` | 字符串（80–13.8k字符） | Opus 4.6完整的思维链推理过程 |
| `solution` | 字符串（20–15.1k字符） | 最终答案或解决方案 |
| `difficulty` | 字符串（3种可选值） | 可选`easy`（简单）、`medium`（中等）或`hard`（困难） |
| `category` | 字符串（2种可选值） | 高级领域分类标签 |
| `timestamp` | 字符串（32字符） | ISO 8601格式的生成时间戳 |
| `hash` | 字符串（16字符） | 去重哈希值 |

---

## 数据集加载

python
from datasets import load_dataset

ds = load_dataset("CompactAI-O/Opus-4.6-Reasoning-2160x", split="train")
print(ds[0]["problem"])
print(ds[0]["thinking"])
print(ds[0]["solution"])

---

## 训练应用场景

本数据集旨在**将前沿推理模型的知识迁移至更小、更高效的模型**，以下为主要训练范式：

### 1. 标准监督微调（Supervised Fine-Tuning，SFT）

以`problem`→`solution`作为训练样本对，用于指令遵循训练。`solution`字段为简洁完整的答案，可直接作为监督目标。

python
def format_sft(row):
return {
"input": row["problem"],
"output": row["solution"]
}

适用场景：已具备基础指令遵循能力，需要迁移领域知识的模型。

### 2. 思维链蒸馏（Chain-of-Thought Distillation）

以`problem`→`thinking + solution`作为训练样本对，用于迁移Opus 4.6的推理风格。`thinking`字段包含了小型/弱模型输出中缺失的多步推理过程。

python
def format_cot(row):
return {
"input": row["problem"],
"output": f"{row['thinking']}

{row['solution']}"
}

适用场景：训练时需要先生成显式推理过程再给出答案的模型（如Qwen、Phi、Mistral以及带有`<|think|>`标记的FANT系列架构）。

### 3. 思维-答案格式（FANT3 / 结构化推理）

针对使用显式推理分隔符的架构（如`<|think|>...<|answer|>...`格式）：

python
def format_think_answer(row, think_open="<|think|>", think_close="<|/think|>",
ans_open="<|answer|>", ans_close="<|/answer|>"):
return (
f"{think_open}{row['thinking']}{think_close}"
f"{ans_open}{row['solution']}{ans_close}"
)

该格式适用于FANT3及同类模型，可在Token层面将隐式推理与最终输出分离。

### 4. 难度加权采样

`difficulty`字段支持课程学习范式——可先从`easy`（简单）样本开始训练，再逐步引入`medium`（中等）与`hard`（困难）样本。

python
from datasets import load_dataset

ds = load_dataset("CompactAI-O/Opus-4.6-Reasoning-2160x", split="train")

easy = ds.filter(lambda x: x["difficulty"] == "easy")
medium = ds.filter(lambda x: x["difficulty"] == "medium")
hard = ds.filter(lambda x: x["difficulty"] == "hard")

### 5. 类别过滤训练

可通过`category`字段将领域特定信号与通用语料混合训练：

python
math_ds = ds.filter(lambda x: x["category"] == "math")
code_ds = ds.filter(lambda x: x["category"] == "code")

### 6. 序列级知识蒸馏（Knowledge Distillation，KD）

搭配本地教师模型使用时，本数据集支持GKD风格的序列级知识蒸馏。`thinking`+`solution`序列可作为教师样本，无需进行Logit对齐，仅需纯序列监督。

python
# Example with TRL SFTTrainer
from trl import SFTTrainer, SFTConfig

trainer = SFTTrainer(
model=student_model,
train_dataset=ds,
args=SFTConfig(
output_dir="./output",
max_seq_length=2048,
),
formatting_func=format_cot,
)
trainer.train()

---

## 任务覆盖领域

样本覆盖以下广泛领域：
- **数学**：代数、几何、算术应用题、组合数学
- **竞赛编程**：字符串处理、图算法、动态规划
- **逻辑与自然语言推理**：前提-假设蕴含、释义检测
- **科学**：化学、生物学、物理学
- **语言任务**：翻译、阅读理解、信息抽取
- **通用问答**：历史、地理、事实性记忆

---

## 数据清洗报告

| 不合格原因 | 数量 | 占初始样本比例 |
|---|---|---|
| 模型拒答/题目不完整 | 1,090 | 33.0% |
| 空响应 | 18 | 0.5% |
| 响应无实质内容 | 17 | 0.5% |
| 题目过短 | 14 | 0.4% |
| 响应过短 | 6 | 0.2% |
| **总计移除** | **1,145** | **34.6%** |
| **总计保留** | **2,160** | **65.4%** |

数据清洗工作于2026年2月12日完成，所有保留样本均具备充实的题目描述与有实质内容的Opus 4.6响应。

---

## 数据去污染

若将本数据集用于评估场景，请先针对你的评估集（如GSM8K、MATH-500、MMLU）进行n-gram去污染处理。原始生成提示中包含的NuminaMath及同类数学相关数据源存在少量重叠风险，与标准基准数据集的逐字污染率估算低于0.5%。

---

## 生成细节

- **模型**：通过OpenRouter调用的Claude Opus 4.6
- **生成时定价**：输入Token每百万Token 5美元，输出Token每百万Token 25美元
- **总输出Token数**：1,636,368
- **估算生成成本**：约409美元
- **平均交互轮次**：1.00（单轮交互，无工具调用）
- **生成日期**：2026年2月

---

## 引用方式

若您使用本数据集，请注明原始生成工作：

@misc{crownelius2026opus46reasoning,
title = {Opus-4.6-Reasoning-2160x},
author = {Crownelius / CompactAI-O},
year = {2026},
url = {https://huggingface.co/datasets/CompactAI-O/Opus-4.6-Reasoning-2160x}
}

---

## 许可证

采用Apache 2.0许可证。Claude Opus 4.6生成的输出需遵守Anthropic的使用政策，对蒸馏后的模型输出进行商业使用时，需符合Anthropic的服务条款。