verifiable-pythonic-function-calling-lite

# Verifiable Pythonic Function Calling Lite

This dataset is a subset of pythonic function calling [dataset](https://huggingface.co/datasets/driaforall/pythonic-function-calling) that is used for training Pythonic function calling models [Dria-Agent-a-3B](https://huggingface.co/driaforall/Dria-Agent-a-3B) and [Dria-Agent-a-7B](https://huggingface.co/driaforall/Dria-Agent-a-7B).

[Dria](docs.dria.co) is a python framework to generate synthetic data on globally connected edge devices with 50+ models. See the network [here](https://dria.co/edge-ai)

## Dataset Summary

The dataset includes various examples of function calling scenarios, ranging from simple to complex multi-turn interactions.
It was generated synthetically using the Dria network on edge devices with various models.

### Categories:
- Simple function calls
- Parallel function executions
- Multiple function interactions
- Multi-turn conversations

## Data

Here's the distribution of example types in the dataset as a Markdown table:

| Type | Description | Percentage |
|------|-------------|------------|
| simple | Single function schema, single function call | 33.5% |
| parallel | Single function schema, multiple function calls | 33.4% |
| multiple | Multiple function schemas, single function call | 22.9% |
| multi_turn | Multiple function schemas, multiple function calls, multi-turn | 10.06% |

Each entry contains:

| Feature Name | Type | Description |
|-------------|------|-------------|
| id | string | Unique identifier for each example |
| domain | string | Main category of the function calling scenario |
| subdomain | string | Specific subcategory within the domain |
| tools | string | Available functions/tools for the scenario in as python functions |
| conversations | list | List of conversation turns |
| - content | string | The actual content of each conversation turn |
| - role | string | The role of the speaker (e.g., user, assistant) |
| type | string | Type of function calling scenario |
| mock_functions | string | Mock implemented function schemas for verification |
| checklist | str (serialzed dict) | A dictionary consisting of "functions" and "values" showing expected function calls and their outputs

> Checklists are json.dump()'ed dictionaries. Load them before you use them.

## Evaluation

We provided a script to evaluate models completions against mock functions and expected values within `evaluate` folder. In order to evaluate, run:
```python
from evaluate.engine import evaluate_row
result = evaluate_row(row["mock_functions"], completion)
```

## License

Apache 2.0

#### Citation

```
@misc{Dria-Agent-a,
url={https://huggingface.co/blog/andthattoo/dria-agent-a},
title={Verifiable Pythonic Function Calling},
author={"andthattoo"}
}
```

# 可验证式Python风格函数调用精简版（Verifiable Pythonic Function Calling Lite）

本数据集为Python风格函数调用（pythonic function calling）数据集（https://huggingface.co/datasets/driaforall/pythonic-function-calling）的子集，用于训练Python风格函数调用模型Dria-Agent-a-3B与Dria-Agent-a-7B。

Dria（docs.dria.co）是一款Python框架，可在全球互联的边缘设备上基于50余种模型生成合成数据，可在此查看其网络架构：https://dria.co/edge-ai

## 数据集摘要
本数据集涵盖各类函数调用场景示例，覆盖从简单到复杂的多轮交互场景，通过Dria网络在边缘设备上使用多款模型合成生成。

### 分类
- 简单函数调用
- 并行函数执行
- 多函数交互
- 多轮对话

## 数据
本数据集的示例类型分布如下Markdown表格所示：

| 类型 | 描述 | 占比 |
|------|-------------|------------|
| simple | 单函数模式（function schema），单函数调用 | 33.5% |
| parallel | 单函数模式（function schema），多函数调用 | 33.4% |
| multiple | 多函数模式（function schema），单函数调用 | 22.9% |
| multi_turn | 多函数模式（function schema）、多函数调用且支持多轮 | 10.06% |

每个数据条目包含以下字段：

| 字段名 | 类型 | 描述 |
|-------------|------|-------------|
| id | 字符串 | 每个示例的唯一标识符 |
| domain | 字符串 | 函数调用场景的主分类 |
| subdomain | 字符串 | 领域内的具体子分类 |
| tools | 字符串 | 场景中可用的函数/工具，以Python函数形式呈现 |
| conversations | 列表 | 对话轮次列表 |
| - content | 字符串 | 单轮对话的实际内容 |
| - role | 字符串 | 发言者角色（如用户、助手） |
| type | 字符串 | 函数调用场景的类型 |
| mock_functions | 字符串 | 用于验证的模拟实现函数模式（function schema） |
| checklist | 字符串（序列化字典） | 包含"functions"与"values"的字典，用于展示预期的函数调用及其输出

> 检查清单为经json.dump()序列化的字典，使用前需先完成反序列化加载。

## 评估
我们在`evaluate`文件夹中提供了用于对照模拟函数与预期值评估模型生成结果的脚本。执行评估的代码如下：
python
from evaluate.engine import evaluate_row
result = evaluate_row(row["mock_functions"], completion)

## 许可证
Apache 2.0

### 引用信息

@misc{Dria-Agent-a,
url={https://huggingface.co/blog/andthattoo/dria-agent-a},
title={可验证式Python风格函数调用},
author={"andthattoo"}
}