AmanPriyanshu/regularizer-250K-from-reasoning-and-tool-use-sft-4M-random-compilation

---
license: cc-by-4.0
---

# Regularizer 250K (from Reasoning + Tool-Use SFT 4M)

250,000 samples combining tool-use agentic data and general reasoning, designed as a regularization set during domain-specific fine-tuning. Preserves tool-use, coding, research, and general reasoning capabilities.

## Construction

1. **Tool-reasoning subset (150K)**: Sampled 50K per category (TOOLS, CODING, RESEARCH) from [AmanPriyanshu/tool-reasoning-sft-1M-random-compilation](https://huggingface.co/datasets/AmanPriyanshu/tool-reasoning-sft-1M-random-compilation) (1M total rows). Random sampling with seed=42.
2. **Reasoning subset (100K)**: Sampled 100K single-turn examples from [AmanPriyanshu/regularizer-250K-from-reasoning-sft-3M-random-compilation](https://huggingface.co/datasets/AmanPriyanshu/regularizer-250K-from-reasoning-sft-3M-random-compilation) (250K total rows).
- Filtered for single-turn only: `len(input) <= 2` (system + user, no prior assistant turns)
- 227,813 single-turn available out of 250K; sampled 100K
- Response parsed: `<think>...</think>` → `reasoning` role, remainder → `answer` role
3. All 250K rows shuffled (seed=42). Messages stored as `json.dumps` strings (use `json.loads` to parse).

## Schema

| Column | Type | Description |
|--------|------|-------------|
| `messages` | `str` (JSON) | JSON string of conversation messages. Use `json.loads()` to parse into `list[{role, content}]` |
| `source_dataset` | `str` | Original source dataset identifier |
| `source_category` | `str` | One of: `TOOLS`, `CODING`, `RESEARCH`, `REASONING` |

## Category Distribution

| Category | Count | % | Source |
|----------|-------|---|--------|
| TOOLS | 50,000 | 20% | tool-reasoning-sft-1M |
| CODING | 50,000 | 20% | tool-reasoning-sft-1M |
| RESEARCH | 50,000 | 20% | tool-reasoning-sft-1M |
| REASONING | 100,000 | 40% | regularizer-250K |
| **Total** | **250,000** | **100%** | |

## Source Dataset Distribution (38 unique)

### TOOLS (50K from 7 sources)

| Source Dataset | Sampled |
|----------------|---------|
| toucan-1.5m-sft-tool-use-data-cleaned-rectified-333k | 9,927 |
| hermes-reasoning-tool-style-data-cleaned-rectified-115k | 9,874 |
| ToolMind-data-cleaned-rectified | 9,747 |
| hermes_reasoning_tool_use-data-cleaned-rectified | 6,599 |
| toolace-sft-tool-use-agent-data-cleaned-rectified | 2,019 |
| mobile-actions-data-cleaned-rectified | 1,131 |
| toolmind-web-qa-sft-tool-use-data-cleaned-rectified-5.2k | 703 |

### CODING (50K from 9 sources)

| Source Dataset | Sampled |
|----------------|---------|
| nvidia-Nemotron-Agentic-v1 | 6,893 |
| Nemotron-Terminal-Corpus-data-cleaned-rectified | 6,881 |
| allenai-SERA-data-cleaned-rectified | 6,858 |
| text_to_terminal_v2-sft-tool-use-agent-data-cleaned-rectified | 6,828 |
| browsing-sft-tool-use-data-cleaned-rectified | 6,807 |
| CoderForge-Preview-data-cleaned-rectified | 6,797 |
| jupyter-agent-dataset-sft-tool-use-agent-data-cleaned-rectified | 6,762 |
| CoVe-12k-data-cleaned-rectified | 1,624 |
| MEnvData-SWE-Trajectory-data-cleaned-rectified | 550 |

### RESEARCH (50K from 8 sources)

| Source Dataset | Sampled |
|----------------|---------|
| OpenHands-CodeScout_Training_Rollouts | 8,960 |
| grill-lab-browsecomp-plus-runs-data-cleaned-rectified | 8,896 |
| explorations | 8,802 |
| rlvr-env-retrieval-source | 8,791 |
| openresearcher-dataset-sft-deep-research-agent-data-cleaned | 8,787 |
| dr-tulu-sft-deep-research-agent-data-cleaned-rectified | 2,392 |
| REDSearcher_SFT_10K | 1,808 |
| OpenSeeker-v1-Data | 1,564 |

### REASONING (100K from 14+ sources)

Sampled from single-turn entries of the [regularizer-250K](https://huggingface.co/datasets/AmanPriyanshu/regularizer-250K-from-reasoning-sft-3M-random-compilation) dataset. Sources include reasoning-sft-CHIMERA, reasoning-sft-OpenThoughts3, reasoning-sft-dolci-think-sft, reasoning-sft-Nemotron-Cascade-SFT-SWE, and others. See parent dataset for full source breakdown.

## Message Format

All rows use a unified `messages` format:

**All rows** use a unified message format with roles: `system`, `user`, `reasoning`, `tool_call`, `tool_output`, `answer`.

**Tool-use rows (TOOLS/CODING/RESEARCH)** — multi-turn agentic:
```json
[
{"role": "system", "content": "..."},
{"role": "user", "content": "..."},
{"role": "reasoning", "content": "<think>...</think>"},
{"role": "tool_call", "content": "..."},
{"role": "tool_output", "content": "..."},
...
{"role": "answer", "content": "..."}
]
```

**Reasoning rows (REASONING)** — single-turn with think tags:
```json
[
{"role": "system", "content": "..."},
{"role": "user", "content": "..."},
{"role": "reasoning", "content": "<think>...</think>"},
{"role": "answer", "content": "..."}
]
```

## Usage

```python
import json
from datasets import load_dataset

ds = load_dataset("AmanPriyanshu/regularizer-250K-from-reasoning-and-tool-use-sft-4M-random-compilation", split="train")
messages = json.loads(ds[0]["messages"]) # list of {role, content} dicts
```

## Purpose

This dataset serves as a **regularizer** during domain-specific fine-tuning (e.g., cybersecurity). By mixing in diverse tool-use and reasoning examples, it prevents catastrophic forgetting of general capabilities while the model specializes.

## Parent Datasets

- [AmanPriyanshu/tool-reasoning-sft-1M-random-compilation](https://huggingface.co/datasets/AmanPriyanshu/tool-reasoning-sft-1M-random-compilation) — 1M multi-turn tool-use SFT samples across TOOLS, CODING, RESEARCH
- [AmanPriyanshu/regularizer-250K-from-reasoning-sft-3M-random-compilation](https://huggingface.co/datasets/AmanPriyanshu/regularizer-250K-from-reasoning-sft-3M-random-compilation) — 250K reasoning samples from the 3M reasoning SFT compilation

license: cc-by-4.0

# 正则化数据集250K（源自4M规模推理+工具使用监督微调数据集）

本数据集包含25万条样本，融合工具使用智能体数据与通用推理数据，旨在作为领域专属微调阶段的正则化数据集，可保留模型的工具使用、代码编写、科研推理与通用推理能力。

## 数据集构建

1. **工具推理子集（15万条）**：从[AmanPriyanshu/tool-reasoning-sft-1M-random-compilation](https://huggingface.co/datasets/AmanPriyanshu/tool-reasoning-sft-1M-random-compilation)（总计100万行数据）中按类别（工具、编码、科研）各采样5万条，采用随机采样策略，随机种子设为42。
2. **推理子集（10万条）**：从[AmanPriyanshu/regularizer-250K-from-reasoning-sft-3M-random-compilation](https://huggingface.co/datasets/AmanPriyanshu/regularizer-250K-from-reasoning-sft-3M-random-compilation)（总计25万行数据）中采样10万条单轮示例。
- 仅筛选单轮示例：`len(input) <= 2`（仅包含系统提示与用户提问，无过往助手回复轮次）
- 原始25万条数据中共计227,813条单轮示例，最终采样10万条
- 响应内容解析规则：将`<think>...</think>`标签内的内容归类为`reasoning`角色，剩余内容归类为`answer`角色
3. 所有25万条数据均采用随机打乱（随机种子为42）。对话消息以`json.dumps`格式的字符串存储（需使用`json.loads`进行解析）。

## 数据结构

| 列名 | 数据类型 | 描述 |
|--------|------|-------------|
| `messages` | `str`（JSON格式） | 对话消息的JSON字符串，需使用`json.loads()`解析为`list[{role, content}]`格式 |
| `source_dataset` | `str` | 原始源数据集标识符 |
| `source_category` | `str` | 可选值为：`TOOLS`、`CODING`、`RESEARCH`、`REASONING` |

## 类别分布

| 类别 | 样本数 | 占比 | 源数据集 |
|----------|-------|---|--------|
| TOOLS | 50,000 | 20% | tool-reasoning-sft-1M |
| CODING | 50,000 | 20% | tool-reasoning-sft-1M |
| RESEARCH | 50,000 | 20% | tool-reasoning-sft-1M |
| REASONING | 100,000 | 40% | regularizer-250K |
| **总计** | **250,000** | **100%** | |

## 源数据集分布（共38个唯一数据集）

### 工具类（5万条，源自7个源数据集）

| 源数据集名称 | 采样数量 |
|----------------|---------|
| toucan-1.5m-sft-tool-use-data-cleaned-rectified-333k | 9,927 |
| hermes-reasoning-tool-style-data-cleaned-rectified-115k | 9,874 |
| ToolMind-data-cleaned-rectified | 9,747 |
| hermes_reasoning_tool_use-data-cleaned-rectified | 6,599 |
| toolace-sft-tool-use-agent-data-cleaned-rectified | 2,019 |
| mobile-actions-data-cleaned-rectified | 1,131 |
| toolmind-web-qa-sft-tool-use-data-cleaned-rectified-5.2k | 703 |

### 编码类（5万条，源自9个源数据集）

| 源数据集名称 | 采样数量 |
|----------------|---------|
| nvidia-Nemotron-Agentic-v1 | 6,893 |
| Nemotron-Terminal-Corpus-data-cleaned-rectified | 6,881 |
| allenai-SERA-data-cleaned-rectified | 6,858 |
| text_to_terminal_v2-sft-tool-use-agent-data-cleaned-rectified | 6,828 |
| browsing-sft-tool-use-data-cleaned-rectified | 6,807 |
| CoderForge-Preview-data-cleaned-rectified | 6,797 |
| jupyter-agent-dataset-sft-tool-use-agent-data-cleaned-rectified | 6,762 |
| CoVe-12k-data-cleaned-rectified | 1,624 |
| MEnvData-SWE-Trajectory-data-cleaned-rectified | 550 |

### 科研类（5万条，源自8个源数据集）

| 源数据集名称 | 采样数量 |
|----------------|---------|
| OpenHands-CodeScout_Training_Rollouts | 8,960 |
| grill-lab-browsecomp-plus-runs-data-cleaned-rectified | 8,896 |
| explorations | 8,802 |
| rlvr-env-retrieval-source | 8,791 |
| openresearcher-dataset-sft-deep-research-agent-data-cleaned | 8,787 |
| dr-tulu-sft-deep-research-agent-data-cleaned-rectified | 2,392 |
| REDSearcher_SFT_10K | 1,808 |
| OpenSeeker-v1-Data | 1,564 |

### 推理类（10万条，源自14+个源数据集）

从[regularizer-250K](https://huggingface.co/datasets/AmanPriyanshu/regularizer-250K-from-reasoning-sft-3M-random-compilation)数据集的单轮条目采样得到，源数据集包括reasoning-sft-CHIMERA、reasoning-sft-OpenThoughts3、reasoning-sft-dolci-think-sft、reasoning-sft-Nemotron-Cascade-SFT-SWE等，完整源数据集明细请参阅父数据集文档。

## 消息格式

所有数据行均采用统一的消息格式，支持以下角色：`system`、`user`、`reasoning`、`tool_call`、`tool_output`、`answer`。

**工具使用类数据（对应工具、编码、科研类别）**为多轮智能体对话格式：
json
[
{"role": "system", "content": "..."},
{"role": "user", "content": "..."},
{"role": "reasoning", "content": "<think>...</think>"},
{"role": "tool_call", "content": "..."},
{"role": "tool_output", "content": "..."},
...
{"role": "answer", "content": "..."}
]

**推理类数据（对应推理类别）**为带有思考标签的单轮对话格式：
json
[
{"role": "system", "content": "..."},
{"role": "user", "content": "..."},
{"role": "reasoning", "content": "<think>...</think>"},
{"role": "answer", "content": "..."}
]

## 使用方法

python
import json
from datasets import load_dataset

ds = load_dataset("AmanPriyanshu/regularizer-250K-from-reasoning-and-tool-use-sft-4M-random-compilation", split="train")
messages = json.loads(ds[0]["messages"]) # 将`messages`字段的JSON字符串解析为对话消息列表

## 数据集用途

本数据集可作为**正则化数据集**应用于领域专属微调阶段（例如网络安全领域微调）。通过混入多样化的工具使用与推理示例，可在模型进行领域专精化微调的同时，避免其出现灾难性遗忘，保留通用能力。

## 父级数据集

- [AmanPriyanshu/tool-reasoning-sft-1M-random-compilation](https://huggingface.co/datasets/AmanPriyanshu/tool-reasoning-sft-1M-random-compilation) —— 涵盖工具、编码、科研类别的100万条多轮工具使用监督微调（Supervised Fine-Tuning，SFT）样本
- [AmanPriyanshu/regularizer-250K-from-reasoning-sft-3M-random-compilation](https://huggingface.co/datasets/AmanPriyanshu/regularizer-250K-from-reasoning-sft-3M-random-compilation) —— 源自300万推理监督微调（Supervised Fine-Tuning，SFT）数据集的25万条推理样本