DJLougen/Acta-Synthetic

---
language:
- en
license: apache-2.0
library_name: datasets
tags:
- agentic
- tool-use
- function-calling
- synthetic
- curated
- high-quality
- conversations
size_categories:
- 1K<n<10K
---

# Acta-Synthetic

High-quality synthetic agentic tool-use conversations generated with 8-factor quality control.

## Overview

This dataset contains **5,000 synthetic agentic tool-use conversations** generated using a template-based synthesis pipeline with strict quality controls.

**Key Statistics:**
- **Train:** 4,500 samples
- **Validation:** 500 samples
- **Avg quality score:** 0.824
- **Avg semantic diversity:** 0.914
- **Avg lexical diversity:** 0.819
- **Pass rate:** 96.2%

## Generation Method

Conversations were generated using the **8-Factor Quality Model**:

| Factor | Weight | Rationale |
|--------|--------|-----------|
| Lexical Diversity | 0.25 | unique_words / total_words (r=+0.967 with quality) |
| Semantic Diversity | 0.20 | content richness (r=+0.947) |
| Verb Uniqueness | 0.15 | unique_verbs / total_verbs (r=+0.852) |
| Turn Efficiency | 0.15 | Optimal 4-10 turns |
| Tool Pattern Novelty | 0.10 | Penalty for over-represented patterns |
| Reasoning Density | 0.08 | Thinking blocks per turn |
| Role Balance | 0.05 | User/assistant ratio |
| Content Brevity | 0.02 | Information density |

**Quality threshold:** 0.75 (only samples scoring above this were kept)

## Domains

- **Database queries:** 15.3%
- **Math problems:** 14.8%
- **API integration:** 12.7%
- **Coding tasks:** 12.3%
- **File processing:** 11.6%
- **Research:** 11.1%
- **Multi-step research:** 11.1%
- **Data analysis:** 11.1%

## Tools Available

- `web_search` - Web search capability
- `wikipedia_search` - Wikipedia knowledge retrieval
- `python_interpreter` - Code execution
- `code_analyzer` - Code analysis and review
- `read_file` - File reading
- `write_file` - File writing
- `api_request` - HTTP API calls
- `database_query` - SQL database queries
- `calculator` - Mathematical calculations
- `data_transform` - Data transformation

## Comparison with Real Data

Validating against [DJLougen/Acta](https://huggingface.co/datasets/DJLougen/Acta) (real agentic data):

| Metric | Real (Acta) | Synthetic | Difference |
|--------|-------------|-----------|------------|
| Quality Score | 0.788 ± 0.052 | 0.826 ± 0.038 | **+0.038** |
| Acceptance Rate | - | 57.2% | - |
| Unique Patterns | - | 234 | - |

**Synthetic data achieves higher quality scores than real curated data** due to:
1. Strict 8-factor filtering during generation
2. Pattern deduplication (max 50 samples per tool pattern)
3. Optimized conversation structure (4-10 turns)
4. Controlled diversity in vocabulary and reasoning

## Usage

```python
from datasets import load_dataset

# Load synthetic dataset
dataset = load_dataset("DJLougen/Acta-Synthetic")

# Access quality metrics
sample = dataset["train"][0]
print(sample["quality_score"]) # 0.824
print(sample["semantic_diversity"]) # 0.914
print(sample["lexical_diversity"]) # 0.819
print(sample["tools_used"]) # ["python_interpreter", "calculator"]
```

## Mixing with Real Data

```python
from datasets import load_dataset
import random

# Load both datasets
real = load_dataset("DJLougen/Acta")["train"]
synthetic = load_dataset("DJLougen/Acta-Synthetic")["train"]

# Mix with 30% synthetic ratio
real_samples = [real[i] for i in range(min(3500, len(real)))]
synthetic_samples = [synthetic[i] for i in range(min(1500, len(synthetic)))]

mixed = real_samples + synthetic_samples
random.shuffle(mixed)

print(f"Mixed dataset: {len(mixed)} samples")
```

## Generation Pipeline

The generation pipeline is available in the [Acta repository](https://huggingface.co/datasets/DJLougen/Acta).

Key components:
- `SyntheticConversationGenerator` - Template-based generation
- `LLMSyntheticGenerator` - LLM-enhanced generation (optional)
- `AgenticDataQualityModel` - 8-factor quality scoring
- `SyntheticDataPipeline` - End-to-end workflow

## Quality Validation

All samples pass strict quality thresholds:
- Lexical diversity ≥ 0.35
- Semantic diversity ≥ 0.40
- Verb uniqueness ≥ 0.01
- Turn count: 2-12

**Target quality score:** ≥ 0.75

## Citation

```bibtex
@dataset{acta_synthetic_2026,
title = {Acta-Synthetic: High-Quality Synthetic Agentic Tool-Use Dataset},
author = {Lougen, Daniel},
year = {2026},
url = {https://huggingface.co/datasets/DJLougen/Acta-Synthetic}
}
```

## License

Apache 2.0

## Related Datasets

- [DJLougen/Acta](https://huggingface.co/datasets/DJLougen/Acta) - Public sample (5K real samples)
- [DJLougen/Acta-Proprietary](https://huggingface.co/datasets/DJLougen/Acta-Proprietary) - Full 37K real samples (private)
- [DJLougen/ornstein-proprietary-100k](https://huggingface.co/datasets/DJLougen/ornstein-proprietary-100k) - 82K samples with 8-factor metrics

---
语言：
- 英语（en）
许可证：Apache 2.0
库名称：datasets
标签：
- 智能体（agentic）
- 工具使用（tool-use）
- 函数调用（function-calling）
- 合成数据（synthetic）
- 精选（curated）
- 高质量（high-quality）
- 对话（conversations）
样本规模类别：
- 1000 < 样本数量 < 10000
---

# Acta-Synthetic

采用八要素质量管控生成的高质量合成智能体工具使用对话集。

## 数据集概览

本数据集包含**5000条合成智能体工具使用对话**，通过基于模板的合成流程结合严格的质量管控生成。

**核心统计数据：**
- 训练集：4500条样本
- 验证集：500条样本
- 平均质量得分：0.824
- 平均语义多样性：0.914
- 平均词汇多样性：0.819
- 通过率：96.2%

## 生成方法

对话通过**八要素质量模型**生成：

| 要素 | 权重 | 原理说明 |
|--------|--------|-----------|
| 词汇多样性（Lexical Diversity） | 0.25 | unique_words / total_words（与质量的相关系数r=+0.967） |
| 语义多样性（Semantic Diversity） | 0.20 | 内容丰富度（相关系数r=+0.947） |
| 动词独特性（Verb Uniqueness） | 0.15 | unique_verbs / total_verbs（相关系数r=+0.852） |
| 对话轮次效率（Turn Efficiency） | 0.15 | 最优轮次为4-10轮 |
| 工具模式新颖性（Tool Pattern Novelty） | 0.10 | 对过度重复的模式施加惩罚 |
| 推理密度（Reasoning Density） | 0.08 | 每轮对话中的思考模块数量 |
| 角色平衡性（Role Balance） | 0.05 | 用户/助手对话占比 |
| 内容简洁性（Content Brevity） | 0.02 | 信息密度 |

**质量阈值：0.75（仅保留得分高于该阈值的样本）**

## 应用领域

- 数据库查询：15.3%
- 数学问题：14.8%
- API集成：12.7%
- 编码任务：12.3%
- 文件处理：11.6%
- 基础研究：11.1%
- 多步研究：11.1%
- 数据分析：11.1%

## 可用工具

- `web_search`：网页搜索功能
- `wikipedia_search`：维基百科知识检索
- `python_interpreter`：代码执行
- `code_analyzer`：代码分析与审查
- `read_file`：文件读取
- `write_file`：文件写入
- `api_request`：HTTP API调用
- `database_query`：SQL数据库查询
- `calculator`：数学计算
- `data_transform`：数据转换

## 与真实数据集对比

基于[DJLougen/Acta](https://huggingface.co/datasets/DJLougen/Acta)（真实智能体对话数据集）进行验证：

| 指标 | 真实数据集（Acta） | 合成数据集 | 差值 |
|--------|-------------|-----------|------------|
| 质量得分 | 0.788 ± 0.052 | 0.826 ± 0.038 | **+0.038** |
| 接受率 | - | 57.2% | - |
| 唯一模式数 | - | 234 | - |

**合成数据集的质量得分高于精选真实数据集，原因如下：**
1. 生成过程中严格执行八要素筛选
2. 模式去重（每个工具模式最多包含50条样本）
3. 优化的对话结构（4-10轮）
4. 词汇与推理多样性可控

## 使用方法

python
from datasets import load_dataset

# 加载合成数据集
dataset = load_dataset("DJLougen/Acta-Synthetic")

# 获取质量指标
sample = dataset["train"][0]
print(sample["quality_score"]) # 0.824
print(sample["semantic_diversity"]) # 0.914
print(sample["lexical_diversity"]) # 0.819
print(sample["tools_used"]) # ["python_interpreter", "calculator"]

## 与真实数据集混合使用

python
from datasets import load_dataset
import random

# 加载两个数据集
real = load_dataset("DJLougen/Acta")["train"]
synthetic = load_dataset("DJLougen/Acta-Synthetic")["train"]

# 按照30%的合成数据占比进行混合
real_samples = [real[i] for i in range(min(3500, len(real)))]
synthetic_samples = [synthetic[i] for i in range(min(1500, len(synthetic)))]

mixed = real_samples + synthetic_samples
random.shuffle(mixed)

print(f"混合后数据集总样本数：{len(mixed)}")

## 生成流程

生成流程可在[Acta仓库](https://huggingface.co/datasets/DJLougen/Acta)中获取。

核心组件包括：
- `SyntheticConversationGenerator`：基于模板的对话生成模块
- `LLMSyntheticGenerator`：大语言模型（LLM）增强生成模块（可选）
- `AgenticDataQualityModel`：八要素质量评分模块
- `SyntheticDataPipeline`：端到端工作流

## 质量验证

所有样本均通过严格的质量阈值校验：
- 词汇多样性 ≥ 0.35
- 语义多样性 ≥ 0.40
- 动词独特性 ≥ 0.01
- 对话轮次：2-12轮

**目标质量得分：≥ 0.75**

## 引用格式

bibtex
@dataset{acta_synthetic_2026,
title = {Acta-Synthetic: High-Quality Synthetic Agentic Tool-Use Dataset},
author = {Lougen, Daniel},
year = {2026},
url = {https://huggingface.co/datasets/DJLougen/Acta-Synthetic}
}

## 许可证

Apache 2.0许可证

## 相关数据集

- [DJLougen/Acta](https://huggingface.co/datasets/DJLougen/Acta)：公开样本集（包含5000条真实样本）
- [DJLougen/Acta-Proprietary](https://huggingface.co/datasets/DJLougen/Acta-Proprietary)：完整37000条真实样本集（私有）
- [DJLougen/ornstein-proprietary-100k](https://huggingface.co/datasets/DJLougen/ornstein-proprietary-100k)：包含82000条带八要素指标的样本集