AmanPriyanshu/reasoning-sft-Edge-Agent-Reasoning-WebSearch-260K

---
license: mit
task_categories:
- text-generation
language:
- en
tags:
- reasoning
- sft
- chain-of-thought
- planning
- agentic
- self-awareness
- web-search
size_categories:
- 100K<n<1M
---

# Edge-Agent Reasoning (Converted)

Converted version of [yatin-superintelligence/Edge-Agent-Reasoning-WebSearch-260K](https://huggingface.co/datasets/yatin-superintelligence/Edge-Agent-Reasoning-WebSearch-260K), reformatted for reasoning SFT training with structured think blocks.

## Format

Each row has three columns:

- **`input`** — list of dicts with system and user messages. System prompt instructs the model to deconstruct problems thoroughly in a `<think>` block, audit its own knowledge, identify ambiguities, and respond with "Planning completed."
- **`response`** — `<think>` block containing the full 2,000-5,000 word reasoning trajectory followed by "Planning completed."
- **`domain`** — industry sector (e.g. Cosmology, Mechanical Engineering, Healthcare, Finance, etc.)

## Reasoning Structure

Each response contains a 5-stage reasoning trajectory:

1. **Understanding the request** — full problem decomposition with constraint identification
2. **What I believe I know — and what I'm uncertain about** — self-aware knowledge auditing with confidence levels
3. **Ambiguities in the request** — identifying missing parameters and conflicting constraints
4. **Everything I need to confirm before responding** — explicit verification checklist
5. **Web search queries** — 10-20 expert-level, keyword-dense queries for RAG retrieval

## Coverage

- 200+ professional roles across 200+ industries
- 39 operating system environments (macOS, Windows, Linux, Android, iPadOS, Cloud Shells)
- Tasks spanning from low difficulty to impossible, safe to catastrophic risk

## License

MIT — inherited from the [original dataset](https://huggingface.co/datasets/yatin-superintelligence/Edge-Agent-Reasoning-WebSearch-260K).

## Credits

Original dataset by [Yatin Taneja](https://huggingface.co/yatin-superintelligence).

许可证：MIT协议
任务类别：
- 文本生成
语言：
- 英语
标签：
- 推理
- 监督微调（Supervised Fine-Tuning, SFT）
- 思维链（Chain-of-Thought）
- 规划
- 智能体式（Agentic）
- 自我意识
- 网页搜索
规模类别：
- 10万条<样本量<100万条

# 边缘智能体推理（转换版）

本数据集为[yatin-superintelligence/Edge-Agent-Reasoning-WebSearch-260K](https://huggingface.co/datasets/yatin-superintelligence/Edge-Agent-Reasoning-WebSearch-260K)的转换版本，经过重新格式化以适配带结构化思考模块的推理监督微调训练。

## 数据格式

每条数据包含三列：

- **`input`**：由包含系统提示与用户消息的字典组成的列表。系统提示要求模型在`<think>`模块中对问题进行彻底拆解，审计自身知识库，识别歧义信息，并最终以"Planning completed."作为回应收尾。
- **`response`**：包含完整2000至5000词推理过程的`<think>`模块，最终以"Planning completed."收尾。
- **`domain`**：所属行业领域（例如宇宙学、机械工程、医疗健康、金融等）

## 推理架构

每条回复包含五阶段推理流程：

1. **理解任务需求**：完整拆解问题并识别约束条件
2. **已知与未知内容**：基于置信度的自我认知知识库审计
3. **任务需求中的歧义点**：识别缺失参数与冲突约束
4. **回复前需确认的全部内容**：明确的验证清单
5. **网页搜索查询**：10至20条具备专家级语义密度的关键词查询，用于检索增强生成（Retrieval-Augmented Generation, RAG）

## 覆盖范围

- 覆盖200余个行业的200余种专业岗位
- 39种操作系统环境（包括macOS、Windows、Linux、Android、iPadOS、云Shell等）
- 任务难度覆盖从低难度到近乎无解，风险等级从安全到灾难性风险

## 许可证

MIT协议——继承自[原始数据集](https://huggingface.co/datasets/yatin-superintelligence/Edge-Agent-Reasoning-WebSearch-260K)。

## 致谢

原始数据集由[Yatin Taneja](https://huggingface.co/yatin-superintelligence)创建。