Implementation:Hiyouga LLaMA Factory Glaive Toolcall Zh Demo Data

Knowledge Sources	Hiyouga_LLaMA_Factory
Domains	NLP, Training_Data
Last Updated	2026-02-06 19:00 GMT

Overview

glaive_toolcall_zh_demo.json provides Chinese tool/function calling demo data in ShareGPT format with tool definitions and multi-turn conversations for training models to invoke external functions in Chinese.

Description

The file contains a JSON array of records, each with a conversations list and a tools string. The conversations follow ShareGPT format using role tags "from" with values "human", "gpt", "function_call", and "observation". All user messages, tool descriptions, tool arguments, and assistant responses are in Simplified Chinese. The "function_call" turns contain JSON-serialized tool invocations, and "observation" turns contain JSON-serialized tool results. The tools field defines available function schemas with Chinese descriptions.

Some conversations demonstrate pure tool invocation flows (e.g., generating invoices, fetching news), while others are purely conversational without tool calls (e.g., explaining programming concepts), training the model to differentiate when tools should be used.

Usage

This demo dataset trains and tests function-calling capabilities for Chinese language models. Users reference it by name (glaive_toolcall_zh_demo) with --stage sft to fine-tune models for Chinese tool use scenarios.

Code Reference

Source Location

• Repository: Hiyouga_LLaMA_Factory
• File: data/glaive_toolcall_zh_demo.json

Data Format

[
  {
    "conversations": [
      {
        "from": "human",
        "value": "我需要为John Doe生成一张发票。他购买了2个苹果，每个$1，以及3根香蕉，每根$0.5。"
      },
      {
        "from": "function_call",
        "value": "{\"name\": \"generate_invoice\", \"arguments\": {\"customer_name\": \"约翰·多伊\", \"items\": [{\"name\": \"苹果\", \"quantity\": 2, \"price\": 1}, {\"name\": \"香蕉\", \"quantity\": 3, \"price\": 0.5}]}}"
      },
      {
        "from": "observation",
        "value": "{\"invoice_id\": \"INV12345\", \"customer_name\": \"约翰·多伊\", \"total\": 3.5, \"status\": \"生成\"}"
      },
      {
        "from": "gpt",
        "value": "发票已成功生成。发票编号为INV12345。约翰·多伊的总金额为$3.5。"
      }
    ],
    "tools": "[{\"name\": \"generate_invoice\", \"description\": \"生成发票\", \"parameters\": {\"type\": \"object\", \"properties\": {\"customer_name\": {\"type\": \"string\", \"description\": \"客户名称\"}, ...}, \"required\": [\"customer_name\", \"items\"]}}]"
  }
]

I/O Contract

Schema

Field	Type	Required	Description
conversations	array	Yes	List of conversation turns in Chinese with "from" and "value" fields
tools	string	Yes	JSON-serialized array of tool/function definitions with Chinese descriptions

Conversation Roles

Role (from)	Description
human	User message in Chinese
gpt	Assistant response in Chinese (text only, no tool call)
function_call	Tool invocation with JSON payload containing name and arguments
observation	Tool execution result as JSON (values often in Chinese)

Dataset Registry Entry

Property	Value
Key	glaive_toolcall_zh_demo
file_name	glaive_toolcall_zh_demo.json
formatting	sharegpt
columns.messages	conversations
columns.tools	tools
Lines	9022

Usage Examples

# Train a model with Chinese function-calling capabilities
# llamafactory-cli train \
#     --dataset glaive_toolcall_zh_demo \
#     --stage sft \
#     --model_name_or_path meta-llama/Llama-2-7b-hf \
#     --output_dir output/toolcall_zh_demo

# Loading and inspecting the data
import json

with open("data/glaive_toolcall_zh_demo.json", "r", encoding="utf-8") as f:
    data = json.load(f)

print(f"Number of conversations: {len(data)}")
sample = data[0]
print(f"Turns in first conversation: {len(sample['conversations'])}")
tools = json.loads(sample['tools'])
print(f"Number of tools defined: {len(tools)}")
for tool in tools:
    print(f"  Tool: {tool['name']} - {tool.get('description', 'N/A')}")

# Count conversations with actual tool calls
tool_call_convs = sum(
    1 for item in data
    if any(t["from"] == "function_call" for t in item["conversations"])
)
print(f"Conversations with tool calls: {tool_call_convs}/{len(data)}")

Related Pages

• Implementation:Hiyouga_LLaMA_Factory_Glaive_Toolcall_En_Demo_Data - English version of the tool-calling demo dataset
• Implementation:Hiyouga_LLaMA_Factory_Dataset_Info_Registry - Central dataset registry that indexes this file
• Implementation:Hiyouga_LLaMA_Factory_Alpaca_Zh_Demo_Data - Chinese SFT demo data (non-tool-calling)