GXMZU/ai-expert-alpaca

AI Expert Alpaca Dataset

🚀 Empower open-source LLMs (Qwen, Gemma, etc.) for core AI domains through SFT/LoRA fine-tuning 🚀

Dataset Description

This dataset contains high-quality Q&A pairs for supervised fine-tuning (SFT) of large language models, focusing on three core AI technology domains: Large Language Models (LLM), Retrieval-Augmented Generation (RAG), and Agent Systems. The dataset provides comprehensive coverage of these cutting-edge AI technologies in both English and Chinese.

Dataset Summary

• Training Set: 11,235 examples in Alpaca format
• Test-AI Set: 1,000 examples focused on LLM, RAG, and Agent technical concepts
• Test-Commonsense Set: 110 examples testing general knowledge and reasoning

Dataset Composition

The training set (11,235 examples) is composed of three categories:

Category	Examples	Percentage	Description
AI Core Technology	9,735	86.65%	LLM, RAG, and Agent domain-specific Q&A pairs
General Instruction	1,000	8.90%	Sampled from vicgalle/alpaca-gpt4
Anti-Hallucination Negatives	500	4.45%	Adversarial samples for hallucination mitigation

AI Core Technology Data (9,735 examples)

• Manually Curated + Model-Assisted: 5,118 examples - High-quality Q&A pairs generated through human curation combined with advanced LLM capabilities
• Semi-Automated Generation: 4,617 examples - Extracted from technical literature using Easy Dataset tool

General Instruction Data (1,000 examples)

Randomly sampled from the open-source vicgalle/alpaca-gpt4 dataset to maintain general language understanding and logical reasoning capabilities, preventing "catastrophic forgetting" during domain-specific fine-tuning.

Anti-Hallucination Negative Samples (500 examples)

Adversarial samples designed to establish clear knowledge boundaries and reduce hallucination. These include four types of traps:

• Fictional Terms: Queries about non-existent algorithms or architectures
• Temporal Confusion: Questions about unreleased or future technologies
• Misattribution: Incorrectly attributing features of one technology to another
• Logical Traps: Questions based on incorrect premises

Dataset Structure

Data Splits

Split	Examples	Description
train	11,235	Training data in Alpaca format
test_ai	1,000	LLM, RAG, Agent technical evaluation set
test_commonsense	110	General knowledge evaluation set

Data Fields

Each split contains different fields appropriate for its purpose:

Training Split (train) - JSONL format

{
  "instruction": "The input prompt or question",
  "input": "Additional context (usually empty)",
  "output": "The expected response"
}

Test-AI Split (test_ai) - JSONL format

{
  "instruction": "The test question focused on LLM, RAG, or Agent technologies",
  "input": "Additional context (usually empty)",
  "expected_output": "Reference answer for evaluation",
  "actual_output": "Field for model predictions (empty by default)"
}

Test-Commonsense Split (test_commonsense) - JSONL format

{
  "Question": "The test question",
  "Reference_Answer": "Reference answer for evaluation",
  "Actual_Output": "Field for model predictions (empty by default)"
}

Note: Each split has its own schema optimized for its specific use case.

Usage

Loading the Dataset

from datasets import load_dataset

# Load training data
dataset = load_dataset("GXMZU/ai-expert-alpaca")
train_data = dataset['train']

# Example: accessing training data
for example in train_data.select(range(5)):
    print(f"Instruction: {example['instruction']}")
    print(f"Output: {example['output']}")
    print("---")

# Test files are available for manual download from the repository:
# - SFT_testdatset_ai.jsonl (1,000 examples)
# - SFT_testdatset_commonsense.jsonl (110 examples)

Licensing

This dataset is released under the MIT License.

Citation

If you use this dataset, please cite:

@dataset{ai_expert_alpaca_2025,
  title={AI Expert Alpaca Dataset},
  author={real-jiakai},
  year={2025},
  month={12},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/GXMZU/ai-expert-alpaca}
}