zake7749/Qwen3.5-27B-DeepCoder-SFT

---
language:
- en
license: cc-by-4.0
task_categories:
- text-generation
tags:
- code
- competitive-programming
- rejection-sampling
- sft
- reasoning
- python
pretty_name: Qwen3.5-27B DeepCoder SFT
size_categories:
- 10K<n<100K
---

## Overview

This dataset contains **14,683 verified code solutions** with chain-of-thought reasoning, generated by `Qwen3.5-27B` on competitive programming problems. Each example has been executed against test cases in a sandboxed environment and passes 100% of tests.

Key features:
- **Reasoning traces** — 77.5% of examples include step-by-step reasoning before the final code solution
- **Verified correctness** — every solution passes all test cases (up to 30 per problem)
- **Rejection sampled** — 8 candidates generated per problem, only passing solutions retained

---

## Dataset Statistics

| Metric | Value |
|--------|-------|
| Total examples | 14,683 |
| Source | DeepCoder (100%) |
| Problems covered | ~4,700 |
| With reasoning | 11,379 (77.5%) |
| Without reasoning (code only) | 3,304 (22.5%) |
| Test cases per problem | mean 29.5, median 30 |

---

## Field Descriptions

| Field | Type | Description |
|-------|------|-------------|
| `id` | string | Unique identifier: `{problem_id}_sft_{candidate_index}` |
| `problem` | string | Full problem statement in English |
| `reasoning` | string | Chain-of-thought reasoning before the solution (empty for 22.5% of examples) |
| `answer` | string | Final answer containing the code block |
| `response` | string | Complete raw model output (reasoning + answer combined, for backward compatibility) |
| `solution` | string | Extracted Python code from the answer |
| `test_pass_rate` | float | Always 1.0 (all tests passed) |
| `tests_passed` | int | Number of test cases passed |
| `tests_total` | int | Total number of test cases executed |
| `source` | string | Data source (`"deepcoder"`) |

---

## Usage

```python
from datasets import load_dataset

ds = load_dataset("zake7749/Qwen3.5-27B-OpenCode-SFT", split="train")

# Only examples with reasoning
with_reasoning = ds.filter(lambda x: len(x["reasoning"]) > 0)
print(f"With reasoning: {len(with_reasoning)}")

# Access fields
example = ds[0]
print(example["reasoning"]) # chain-of-thought
print(example["solution"]) # extracted Python code
```

---

## Related Datasets

- [zake7749/Qwen3-Coder-Next-OpenCode-SFT](https://huggingface.co/datasets/zake7749/Qwen3-Coder-Next-OpenCode-SFT) — 49,374 SFT examples from Qwen3-Coder-Next
- [zake7749/Qwen3-Coder-Next-OpenCode-Preference](https://huggingface.co/datasets/zake7749/Qwen3-Coder-Next-OpenCode-Preference) — 10,920 preference pairs for DPO/KTO

## Call for Collaboration

We are actively looking for **compute sponsors and research collaborators** to scale up this work. Our roadmap includes:

- **Scaling data generation** — more problems, more candidates per problem, and broader language coverage
- **SFT & RL training** — training and open-sourcing code reasoning models using this dataset and its SFT counterpart
- **Expanding the pipeline** — supporting more base models beyond Qwen3-Coder-Next

We welcome all forms of collaboration:
- **Compute sponsorship** — GPU hours for large-scale generation and training runs
- **Joint research** — co-developing models, co-authoring technical reports

## License

Source data licenses should be verified before redistribution:
- **DeepCoder** (`agentica-org/DeepCoder-Preview-Dataset`): Check HuggingFace dataset card

## Citation

```bibtex
@misc{yang_qwen35_opencode_sft_2026,
title={Qwen3.5-27B-OpenCode-SFT},
author={Yang, Kai-Chou},
year={2026},
publisher={Hugging Face},
url={https://huggingface.co/datasets/zake7749/Qwen3.5-27B-OpenCode-SFT/}
}
```

---
language:
- 英语
license: CC BY 4.0
task_categories:
- 文本生成
tags:
- 代码
- 编程竞赛
- 拒绝采样
- 监督微调（Supervised Fine-Tuning，SFT）
- 推理
- Python
pretty_name: Qwen3.5-27B DeepCoder 监督微调数据集
size_categories:
- 10000 < 样本数 < 100000
---

## 数据集概览

本数据集包含**14683份经过验证的代码解决方案**，附带思维链（Chain-of-Thought，CoT）推理过程，由`Qwen3.5-27B`针对编程竞赛题目生成。所有样本均在沙箱环境中针对测试用例执行，且100%通过所有测试。

核心特性：
- **推理轨迹**：77.5%的样本在最终代码解决方案前包含逐步推理过程
- **正确性验证**：每份解决方案均通过所有测试用例（单题最多包含30个测试用例）
- **拒绝采样**：每道题目生成8个候选方案，仅保留通过测试的解决方案

## 数据集统计信息

| 指标 | 数值 |
| ---- | ---- |
| 总样本数 | 14683 |
| 数据来源 | DeepCoder（100%） |
| 覆盖题目数 | 约4700道 |
| 带推理过程样本 | 11379份（占比77.5%） |
| 无推理过程样本（仅含代码） | 3304份（占比22.5%） |
| 单题测试用例数 | 均值29.5，中位数30 |

## 字段说明

| 字段 | 数据类型 | 描述 |
| ---- | ---- | ---- |
| `id` | 字符串 | 唯一标识符，格式为`{problem_id}_sft_{candidate_index}` |
| `problem` | 字符串 | 英文完整题目描述 |
| `reasoning` | 字符串 | 解决方案前的思维链推理过程，22.5%的样本此字段为空 |
| `answer` | 字符串 | 包含代码块的最终答案 |
| `response` | 字符串 | 模型完整原始输出（推理过程与答案的组合，用于向后兼容） |
| `solution` | 字符串 | 从答案中提取的Python代码 |
| `test_pass_rate` | 浮点数 | 固定为1.0（所有测试均通过） |
| `tests_passed` | 整数 | 通过的测试用例数量 |
| `tests_total` | 整数 | 执行的测试用例总数 |
| `source` | 字符串 | 数据来源，固定为`"deepcoder"` |

## 使用方法

python
from datasets import load_dataset

ds = load_dataset("zake7749/Qwen3.5-27B-OpenCode-SFT", split="train")

# 仅筛选包含推理过程的样本
with_reasoning = ds.filter(lambda x: len(x["reasoning"]) > 0)
print(f"包含推理过程的样本数：{len(with_reasoning)}")

# 访问字段内容
example = ds[0]
print(example["reasoning"]) # 思维链推理内容
print(example["solution"]) # 提取的Python代码

## 相关数据集

- [zake7749/Qwen3-Coder-Next-OpenCode-SFT](https://huggingface.co/datasets/zake7749/Qwen3-Coder-Next-OpenCode-SFT)：来自Qwen3-Coder-Next的49374条监督微调（SFT）样本
- [zake7749/Qwen3-Coder-Next-OpenCode-Preference](https://huggingface.co/datasets/zake7749/Qwen3-Coder-Next-OpenCode-Preference)：用于直接偏好优化（Direct Preference Optimization，DPO）与知识增强时序优化（Knowledge-based Temporal Optimization，KTO）的10920条偏好配对样本

## 合作招募

我们正积极寻求**算力赞助方与研究合作者**，以推进该项目的规模化发展。我们的路线图包含：
- **扩大数据生成规模**：覆盖更多题目、增加单题候选方案数量、拓展语言覆盖范围
- **监督微调与强化学习训练**：使用本数据集及其配套监督微调数据集训练并开源代码推理模型
- **扩展流程管线**：支持Qwen3-Coder-Next以外的更多基础模型

我们欢迎各类形式的合作：
- **算力赞助**：为大规模生成与训练任务提供GPU算力时长
- **联合研究**：共同开发模型、共同撰写技术报告

## 许可证

重新分发本数据集前，请验证源数据的许可证：
- **DeepCoder**（`agentica-org/DeepCoder-Preview-Dataset`）：请查看Hugging Face数据集卡片获取详细信息

## 引用格式

bibtex
@misc{yang_qwen35_opencode_sft_2026,
title={Qwen3.5-27B-OpenCode-SFT},
author={Yang, Kai-Chou},
year={2026},
publisher={Hugging Face},
url={https://huggingface.co/datasets/zake7749/Qwen3.5-27B-OpenCode-SFT/}
}