Implementation:Hiyouga LLaMA Factory Alpaca Zh Demo Data

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

Overview

alpaca_zh_demo.json provides 500 Chinese instruction-response pairs in the Alpaca format for demonstrating and testing supervised fine-tuning (SFT) workflows in LLaMA Factory.

Description

The file contains a JSON array of 500 records in Chinese, each following the standard Alpaca data format with three fields: instruction, input, and output. The instructions cover diverse topics such as scientific explanations, programming tasks, classification, creative writing, and general knowledge questions -- all in Simplified Chinese. The input field is optionally used when the task requires additional context, and the output field contains the expected Chinese model response.

This dataset uses the default Alpaca formatting and is registered in dataset_info.json under the key "alpaca_zh_demo" with only a file_name property, requiring no special column mappings.

Usage

This demo dataset is intended for quick testing and validation of SFT training pipelines for Chinese language models. Users reference it by name (alpaca_zh_demo) in their training configuration to verify data loading and training processes work correctly with Chinese text before using larger production datasets.

Code Reference

Source Location

Repository: Hiyouga_LLaMA_Factory
File: data/alpaca_zh_demo.json

Data Format

[
  {
    "instruction": "识别并解释给定列表中的两个科学理论：细胞理论和日心说。",
    "input": "",
    "output": "细胞理论是生物科学的一个理论，它认为所有生命体都是由微小的基本单元——细胞所构成..."
  },
  {
    "instruction": "分类以下数字系列。",
    "input": "输入：\n2、4、6、8",
    "output": "此系列数字为偶数序列。"
  }
]

I/O Contract

Schema

Field	Type	Required	Description
instruction	string	Yes	The task or question in Chinese for the model to respond to
input	string	No	Additional context or input data in Chinese (empty string if unused)
output	string	Yes	The expected model response in Chinese

Dataset Registry Entry

Property	Value
Key	`alpaca_zh_demo`
file_name	`alpaca_zh_demo.json`
formatting	alpaca (default)
Lines	5002
Records	~500

Usage Examples

# Reference the dataset in a LLaMA Factory YAML training config
# examples/train_lora/llama3_lora_sft.yaml
#   dataset: alpaca_zh_demo

# Or use it directly via CLI
# llamafactory-cli train \
#     --dataset alpaca_zh_demo \
#     --stage sft \
#     --model_name_or_path meta-llama/Llama-2-7b-hf \
#     --output_dir output/sft_zh_demo

# Loading the data manually for inspection
import json

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

print(f"Number of samples: {len(data)}")
print(f"First sample instruction: {data[0]['instruction'][:40]}...")
print(f"Has input: {bool(data[0]['input'])}")

Related Pages

Hiyouga_LLaMA_Factory_Alpaca_En_Demo_Data - English version of the Alpaca demo dataset
Hiyouga_LLaMA_Factory_Dataset_Info_Registry - Central dataset registry that indexes this file
Hiyouga_LLaMA_Factory_Identity_Data - Identity training data using the same Alpaca format

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment