Weni/Function-Calling-Benchmark-1.1.0

# Dataset Card – Function Calling SFT (Bilingual)

**Dataset Name:** Weni/Function-Calling-Benchmark-1.1.0

## Description
This dataset was created for **Supervised Fine-Tuning (SFT)** of language models designed to perform **structured function calling**.
It provides bilingual examples (Portuguese and English) that pair natural language inputs with structured tool invocation outputs, helping models learn to reason over parameters and select the appropriate function (“tool”) to call.

Each instance represents an instruction where the model must:
1. Interpret the input query;
2. Select the correct tool from a provided list (`Tools`);
3. Generate the correct structured output under the `parameters` field, now represented as a **list of dictionaries** for each tool call.

The dataset was designed to improve **tool-use accuracy**, **parameter formatting**, and **cross-language consistency** in models trained for API-based reasoning or agentic tasks.

---

## Dataset Attributes

| Field | Type | Description | Values / Notes |
|-------|------|-------------|----------------|
| **Tools** | list[dict] | List of available tools for the model to choose from, each containing `tool` (tool name) and `description` (tool context). | Example: `[{"tool": "get_weather", "description": "Returns current weather data for a given city"}]` |
| **tool_selected** | int | Index of the tool chosen by the model for this example. | 0-based index |
| **parameters** | list[dict] | Parameters passed to the selected tool. Stored as a list of dictionaries to preserve structure and type information. | Example: `[{"city": "Florianópolis", "unit": "Celsius"}]` |
| **user_input** | string / text | The user’s natural-language instruction or query that triggers the function call. | Example: “Show me the weather in São Paulo in Celsius.” |
| **language** | integer (enum) | Language indicator for bilingual SFT alignment. | `0 = Portuguese`, `1 = English` |

---

## Usage and Application
This dataset is ideal for training or evaluating **LLMs with function-calling capabilities**, particularly in **bilingual** contexts (English–Portuguese).
It can be used to:

- Fine-tune models on **tool-selection and parameter-inference tasks**;
- Benchmark **API call generation** consistency across languages;
- Evaluate structured reasoning in **agentic or orchestration pipelines** (e.g., assistants that decide which function to call and how to format arguments).

---

**Recommended Use Cases**
- Fine-tuning multilingual assistant models with OpenAI-style function-calling schemas.
- Training bilingual agent frameworks (LangChain, Bedrock Agents, etc.).
- Evaluating structured reasoning and argument serialization in LLMs.