DeepSeek-V4-Distill-8000x

# 🐳 DeepSeek-V4-Distill-8100x

<!-- markdownlint-disable first-line-h1 -->
<!-- markdownlint-disable html -->
<!-- markdownlint-disable no-duplicate-header -->

<div align="center">
<img src="https://github.com/deepseek-ai/DeepSeek-V2/blob/main/figures/logo.svg?raw=true" width="60%" alt="DeepSeek-V4" />
</div>


## Dataset Summary

DeepSeek-V4-Distill-8100x is a supervised fine-tuning dataset for reasoning-oriented distillation. The question prompts come from [Jackrong/GLM-5.1-Reasoning-1M-Cleaned](https://huggingface.co/datasets/Jackrong/GLM-5.1-Reasoning-1M-Cleaned), and the answers were generated by the teacher model **DeepSeek-V4-Flash**.

After the cleaning process, the released `train` split contains **7,716** high-quality JSONL examples.

> [!NOTE]
> The answer pool was cleaned to remove real-time questions, identity-related questions, overlong questions, and other unsuitable prompts. This filtering is intended to make the distillation set more stable and reduce prompts that depend on transient facts or model identity behavior.

<div align="center">
<img src="assets/dsv4_performance.png" >
</div>

## Dataset Details

| Item | Value |
| --- | --- |
| Repository | `Jackrong/DeepSeek-V4-Distill-8000x` |
| Source questions | `Jackrong/GLM-5.1-Reasoning-1M-Cleaned` |
| Teacher model | `DeepSeek-V4-Flash` |
| Format | JSONL |
| Split | `train` |
| Examples | 7,716 |
| Language | Primarily English |

## Data Structure

Each row contains both conversation-style and direct input/output fields:

| Field | Type | Description |
| --- | --- | --- |
| `id` | string | Stable sample identifier. |
| `conversations` | list | Conversation with `human` prompt and `gpt` response. |
| `input` | string | User question or instruction. |
| `output` | string | Teacher-generated reasoning response, often including a `<think>...</think>` block followed by the final answer. |
| `domain` | string | Dataset subset label. All current rows use `main`. |
| `meta` | object | Generation metadata, including `input_tokens`, `output_tokens`, and `teacher_model`. |

Example:

```json
{
"id": "84677a95a9ac13cced90a51b38004e62",
"conversations": [
{"from": "human", "value": "..."},
{"from": "gpt", "value": "<think>...</think>\\n\\n..."}
],
"input": "...",
"output": "<think>...</think>\\n\\n...",
"domain": "main",
"meta": {
"input_tokens": 65,
"output_tokens": 95,
"teacher_model": "DeepSeek-V4-Flash"
}
}
```

## Intended Uses

- Reasoning-oriented supervised fine-tuning.
- Distillation experiments using `DeepSeek-V4-Flash` teacher outputs.
- Format conversion experiments for chat-style and input/output SFT pipelines.

## Limitations

- Responses are teacher-generated and may contain factual errors, reasoning artifacts, or style biases inherited from the teacher model.
- The dataset is not intended for safety-critical use without further review.
- Downstream users should check the license and usage terms of upstream sources and generated content for their own use case.

## Statistics



# 🐳 DeepSeek-V4-Distill-8100x

<!-- markdownlint-disable first-line-h1 -->
<!-- markdownlint-disable html -->
<!-- markdownlint-disable no-duplicate-header -->

<div align="center">
<img src="https://github.com/deepseek-ai/DeepSeek-V2/blob/main/figures/logo.svg?raw=true" width="60%" alt="DeepSeek-V4" />
</div>


## 数据集概述

DeepSeek-V4-Distill-8100x 是一款面向推理蒸馏的监督微调数据集。其问题提示源自数据集 [Jackrong/GLM-5.1-Reasoning-1M-Cleaned](https://huggingface.co/datasets/Jackrong/GLM-5.1-Reasoning-1M-Cleaned)，答案由教师模型 **DeepSeek-V4-Flash** 生成。

经过清洗流程后，本次发布的`train`训练划分集包含**7716**条高质量JSONL格式样本。

> [!注意]
> 本次答案池已完成清洗，移除了实时性问题、身份相关问题、过长问题及其他不合适的提示词。该筛选流程旨在提升蒸馏数据集的稳定性，减少依赖瞬时事实或模型身份行为的提示词。

<div align="center">
<img src="assets/dsv4_performance.png" >
</div>

## 数据集详情

| 项目 | 内容 |
| --- | --- |
| 仓库地址 | `Jackrong/DeepSeek-V4-Distill-8000x` |
| 源问题来源 | `Jackrong/GLM-5.1-Reasoning-1M-Cleaned` |
| 教师模型 | `DeepSeek-V4-Flash` |
| 数据格式 | JSONL |
| 划分集 | `train` |
| 样本数量 | 7716 |
| 语言 | 以英语为主 |

## 数据结构

每条样本同时包含对话式与直接输入输出两类字段：

| 字段 | 类型 | 描述 |
| --- | --- | --- |
| `id` | 字符串 | 稳定的样本唯一标识符 |
| `conversations` | 列表 | 包含`human`（用户）提示与`gpt`（助手）回复的对话内容 |
| `input` | 字符串 | 用户的问题或指令 |
| `output` | 字符串 | 教师模型生成的推理回复，通常包含`<think>...</think>`推理块与后续的最终答案 |
| `domain` | 字符串 | 数据集子集标签，当前所有样本均使用`main` |
| `meta` | 对象 | 生成元数据，包含`input_tokens`（输入令牌 (Token) 数）、`output_tokens`（输出令牌 (Token) 数）与`teacher_model`（教师模型） |

示例：

json
{
"id": "84677a95a9ac13cced90a51b38004e62",
"conversations": [
{"from": "human", "value": "..."},
{"from": "gpt", "value": "<think>...</think>\n\n..."}
],
"input": "...",
"output": "<think>...</think>\n\n...",
"domain": "main",
"meta": {
"input_tokens": 65,
"output_tokens": 95,
"teacher_model": "DeepSeek-V4-Flash"
}
}

## 预期用途

- 面向推理的监督微调任务
- 基于`DeepSeek-V4-Flash`教师模型输出的蒸馏实验
- 对话式与输入输出式监督微调（SFT）流水线的格式转换实验

## 局限性

- 回复由教师模型生成，可能包含事实错误、推理瑕疵或继承自教师模型的风格偏见
- 未经额外审核的情况下，本数据集不适合用于安全敏感型应用
- 下游使用者需针对自身使用场景，核查上游数据源与生成内容的许可协议及使用条款

## 统计信息