DJLougen/Acta

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

# Acta

*Acta* (Latin: "acts, deeds, records") - A premium curated sample of high-quality agentic tool-use conversations, filtered using an 8-factor quality model based on statistical correlation analysis of diversity metrics.

## Overview

This public sample contains **5,000 conversations** from the full 37K Acta dataset. It was created by analyzing 5 major agentic datasets (FlameFox, Hermes Reasoning, Hermes Multi-turn, Smolagents, SWE Agent) and applying evidence-based quality filters to maximize training signal density.

**Key Statistics:**
- **Train:** 4,500 samples
- **Validation:** 500 samples
- **Avg semantic diversity:** 0.53
- **Avg lexical diversity:** 0.47
- **Sources:** 5 datasets harmonized

## The 8-Factor Quality Model

Each sample was scored using 8 factors weighted by their correlation with training quality:

| Factor | Weight | Threshold | Rationale |
|--------|--------|-----------|-----------|
| Lexical Diversity | 0.25 | >=0.35 | unique_words / total_words (r=+0.967 with quality) |
| Semantic Diversity | 0.20 | >=0.40 | content richness excluding stopwords (r=+0.947) |
| Verb Uniqueness | 0.15 | >=0.01 | unique_verbs / total_verbs (r=+0.852) |
| Turn Efficiency | 0.15 | 2-12 turns | Optimal conversation length (longer = repetitive) |
| Tool Pattern Novelty | 0.10 | - | Penalty for over-represented tool sequences |
| Reasoning Density | 0.08 | - | Thinking blocks per assistant turn |
| Role Balance | 0.05 | - | User/assistant turn ratio |
| Content Brevity | 0.02 | - | Information density (shorter = denser signal) |

## Source Distribution

| Source | Samples | Avg Quality | Characteristics |
|--------|---------|-------------|-----------------|
| FlameFox Agentic | ~2,600 | 0.819 | Highest diversity, balanced tools |
| Hermes Reasoning | ~1,400 | 0.598 | Good semantic diversity |
| Hermes Multi-turn | ~450 | 0.534 | Multi-turn, deduplicated |
| Smolagents Code | ~70 | 0.592 | Low redundancy |
| SWE Agent GLM | ~20 | 0.433 | Shortest, least repetitive traces |

## Usage

```python
from datasets import load_dataset

# Load the curated sample
dataset = load_dataset("DJLougen/Acta")

# Access quality metrics
sample = dataset["train"][0]
print(sample["quality_score"]) # 0.0 - 1.0
print(sample["semantic_diversity"]) # 0.546
print(sample["lexical_diversity"]) # 0.400
```

## Key Findings

1. **Quality ≠ Quantity** - 37K curated samples > 200K raw samples
2. **Lexical diversity is the strongest quality predictor** (r=+0.967)
3. **More content ≠ better signal** - SWE Agent has 5x more text but 3x lower quality
4. **4-10 turns optimal** - longer conversations become repetitive (r=-0.982)
5. **Tool redundancy is rampant** - some patterns repeat 3000+ times in raw data

## Citation

```bibtex
@dataset{acta_2026,
title = {Acta: A Quality-Curated Agentic Tool-Use Dataset},
author = {Lougen, Daniel},
year = {2026},
url = {https://huggingface.co/datasets/DJLougen/Acta}
}
```

## License

Apache 2.0

## Acknowledgments

Source datasets:
- [FlameF0X/agentic-code](https://huggingface.co/datasets/FlameF0X/agentic-code)
- [interstellarninja/hermes_reasoning_tool_use](https://huggingface.co/datasets/interstellarninja/hermes_reasoning_tool_use)
- [interstellarninja/tool-use-multiturn-reasoning](https://huggingface.co/datasets/interstellarninja/tool-use-multiturn-reasoning)
- [smolagents/codeagent-traces](https://huggingface.co/datasets/smolagents/codeagent-traces)
- [DCAgent/neulab-nebius-swe-agent-trajectories-sandboxes_glm_4.7_traces_jupiter](https://huggingface.co/datasets/DCAgent/neulab-nebius-swe-agent-trajectories-sandboxes_glm_4.7_traces_jupiter)

---

**Full 37K proprietary dataset:** [DJLougen/Acta-Proprietary](https://huggingface.co/datasets/DJLougen/Acta-Proprietary)

language:
- 英语
license: apache-2.0
许可证：Apache 2.0
library_name: datasets
库名称：datasets
tags:
- 智能体相关（agentic）
- 工具使用（tool-use）
- 函数调用（function-calling）
- 精选（curated）
- 高质量（high-quality）
- 对话（conversations）
size_categories:
- 1K<n<10K
规模类别：1千 < n < 1万

# Acta
*Acta*（拉丁语意为“行动、事迹、记录”）是一份经过精选的高质量智能体工具使用对话样本集，基于对多样性指标的统计相关性分析，通过八因子质量模型完成筛选。

## 概述
本公开样本集包含完整37K条Acta数据集中的**5000条对话**。该数据集通过分析5个主流智能体数据集（FlameFox、Hermes Reasoning、Hermes Multi-turn、Smolagents、SWE Agent），并应用循证质量过滤方法以最大化训练信号密度而构建。

**核心统计数据：**
- **训练集：** 4500个样本
- **验证集：** 500个样本
- **平均语义多样性：** 0.53
- **平均词汇多样性：** 0.47
- **数据源：** 5个经过对齐的数据集

## 八因子质量模型
每个样本均基于其与训练质量的相关性赋予权重，通过以下8个因子进行评分：

| 因子 | 权重 | 阈值 | 说明 |
|--------|--------|-----------|-----------|
| 词汇多样性（Lexical Diversity） | 0.25 | ≥0.35 | 唯一词数/总词数（与质量的相关系数*r*=+0.967） |
| 语义多样性（Semantic Diversity） | 0.20 | ≥0.40 | 去除停用词后的内容丰富度（相关系数*r*=+0.947） |
| 动词独特性（Verb Uniqueness） | 0.15 | ≥0.01 | 唯一动词数/总动词数（相关系数*r*=+0.852） |
| 轮次效率（Turn Efficiency） | 0.15 | 2-12轮 | 最优对话长度（过长则易出现重复内容） |
| 工具模式新颖性（Tool Pattern Novelty） | 0.10 | - | 对过度重复的工具序列施加惩罚 |
| 推理密度（Reasoning Density） | 0.08 | - | 每轮助手回复中的思考块数量 |
| 角色平衡性（Role Balance） | 0.05 | - | 用户与助手的轮次比例 |
| 内容简洁性（Content Brevity） | 0.02 | - | 信息密度（内容越简短则信号密度越高） |

## 数据源分布
| 数据源 | 样本数量 | 平均质量得分 | 特性 |
|--------|---------|-------------|-----------------|
| FlameFox Agentic | ~2600 | 0.819 | 多样性最高，工具使用均衡 |
| Hermes Reasoning | ~1400 | 0.598 | 语义多样性表现优异 |
| Hermes Multi-turn | ~450 | 0.534 | 多轮对话，已去重 |
| Smolagents Code | ~70 | 0.592 | 冗余度极低 |
| SWE Agent GLM | ~20 | 0.433 | 篇幅最短，重复痕迹最少 |

## 使用方法
python
from datasets import load_dataset

# 加载精选样本集
dataset = load_dataset("DJLougen/Acta")

# 访问质量指标
sample = dataset["train"][0]
print(sample["quality_score"]) # 取值范围0.0 - 1.0
print(sample["semantic_diversity"]) # 0.546
print(sample["lexical_diversity"]) # 0.400

## 核心发现
1. **质量≠数量**：37K条精选样本优于200K条原始样本
2. **词汇多样性是最强的质量预测因子**（相关系数*r*=+0.967）
3. **内容更多≠信号更佳**：SWE Agent数据集的文本量是其他数据集的5倍，但质量仅为其三分之一
4. **4-10轮为最优对话长度**：过长的对话易出现重复（相关系数*r*=-0.982）
5. **工具使用模式重复现象严重**：原始数据中部分模式重复次数超过3000次

## 引用格式
bibtex
@dataset{acta_2026,
title = {Acta: 经质量精选的智能体工具使用数据集},
author = {Lougen, Daniel},
year = {2026},
url = {https://huggingface.co/datasets/DJLougen/Acta}
}

## 许可证
Apache 2.0

## 致谢
数据源包括：
- [FlameF0X/agentic-code](https://huggingface.co/datasets/FlameF0X/agentic-code)
- [interstellarninja/hermes_reasoning_tool_use](https://huggingface.co/datasets/interstellarninja/hermes_reasoning_tool_use)
- [interstellarninja/tool-use-multiturn-reasoning](https://huggingface.co/datasets/interstellarninja/tool-use-multiturn-reasoning)
- [smolagents/codeagent-traces](https://huggingface.co/datasets/smolagents/codeagent-traces)
- [DCAgent/neulab-nebius-swe-agent-trajectories-sandboxes_glm_4.7_traces_jupiter](https://huggingface.co/datasets/DCAgent/neulab-nebius-swe-agent-trajectories-sandboxes_glm_4.7_traces_jupiter)

---

**完整37K条专有数据集：** [DJLougen/Acta-Proprietary](https://huggingface.co/datasets/DJLougen/Acta-Proprietary)