New announcement. Learn more


Unlocking AI Potential: The Critical Role of High-Quality Data in Large Language Models

High-quality data for AI models

In the dynamic field of artificial intelligence, large language models (LLMs) like OpenAI's GPT-4 and Google's PaLM have revolutionised natural language processing tasks. These models, capable of generating human-like text, performing complex analyses, and engaging in conversations, owe their effectiveness to the quality of the data they are trained on.

Ensuring high-quality data is crucial for maximising the potential of these models and avoiding significant pitfalls.

The Importance of High-Quality Data

The old adage "garbage in, garbage out" holds particularly true for LLMs. High-quality data ensures that these models produce accurate, reliable, and relevant outputs. Conversely, low-quality data can lead to biased, inaccurate, and even harmful results.

Ensuring Accuracy and Reliability

High-quality data is essential for maintaining the accuracy and reliability of LLM outputs. As highlighted by the Reserve Bank of New Zealand, frequent and timely releases of high-quality data are fundamental for sound decision-making and meaningful research.

This principle is equally vital in the realm of AI, where the precision of data directly impacts the model’s predictions and insights​ (Reserve Bank of New Zealand)​​ (Scoop)​.

Avoiding Bias and Errors

Training LLMs on vast amounts of data without considering quality can introduce biases and errors. For instance, generative AI tools, if not properly vetted, can perpetuate existing biases present in the data. Health New Zealand advises a cautious approach due to risks such as privacy breaches, inaccuracies, and biases in the output of generative AI tools​ (Te Whatu Ora)​.

High-quality, diverse data helps mitigate these issues, ensuring more balanced and fair AI applications.

Enhancing Model Performance

The performance of LLMs is closely tied to the quality of the training data. High-quality data helps in developing models that are adaptable and effective across various tasks and scenarios.

The University of Auckland emphasises that integrating high-quality semantic data is crucial for creating reliable and credible digital models, which can be applied to LLMs as well​ (The University of Auckland)​.

Best Practices for High-Quality Data Collection

To harness the full benefits of LLMs, it is crucial to adhere to best practices in data collection and preprocessing. This includes:

Data Deduplication

  • Removing duplicate entries at multiple levels (sentence, document, dataset) to prevent repetitive patterns in model outputs. This helps improve the learning process and ensures cleaner data​ (Education in New Zealand)​.

Privacy Redaction

  • Ensuring that personally identifiable information (PII) is removed from the training data to protect user privacy and comply with regulations​


  • Using customised tokenisers tailored to the specific characteristics of the pre-training corpus enhances the model’s understanding and processing of diverse text formats and languages​​.

Balancing Data Sources

  • Incorporating a mix of data from various domains improves the model’s generalisation ability while avoiding over-reliance on any single data type or source​ (The University of Auckland)​.

Investing in high-quality data is a foundational step in unlocking the full potential of large language models. By prioritising data quality through rigorous preprocessing and careful selection of training datasets, organisations can develop more accurate, reliable, and versatile AI applications.

As AI continues to evolve, maintaining a steadfast commitment to high-quality data will be essential for achieving sustainable success and innovation.

By incorporating insights and best practices from leading New Zealand institutions and experts, this article underscores the critical role of high-quality data in AI and LLMs.

For more detailed information, you can refer to the resources from the Reserve Bank of New Zealand, Health New Zealand, and the University of Auckland.