Towards Best Practices for Open Datasets for LLM Training (2501.08365v1)

Abstract: Many AI companies are training their LLMs on data without the permission of the copyright owners. The permissibility of doing so varies by jurisdiction: in countries like the EU and Japan, this is allowed under certain restrictions, while in the United States, the legal landscape is more ambiguous. Regardless of the legal status, concerns from creative producers have led to several high-profile copyright lawsuits, and the threat of litigation is commonly cited as a reason for the recent trend towards minimizing the information shared about training datasets by both corporate and public interest actors. This trend in limiting data information causes harm by hindering transparency, accountability, and innovation in the broader ecosystem by denying researchers, auditors, and impacted individuals access to the information needed to understand AI models. While this could be mitigated by training LLMs on open access and public domain data, at the time of writing, there are no such models (trained at a meaningful scale) due to the substantial technical and sociological challenges in assembling the necessary corpus. These challenges include incomplete and unreliable metadata, the cost and complexity of digitizing physical records, and the diverse set of legal and technical skills required to ensure relevance and responsibility in a quickly changing landscape. Building towards a future where AI systems can be trained on openly licensed data that is responsibly curated and governed requires collaboration across legal, technical, and policy domains, along with investments in metadata standards, digitization, and fostering a culture of openness.

• The paper establishes normative principles emphasizing transparency and reproducibility to foster a competitive open dataset ecosystem for LLM training.
• It details practical recommendations for sourcing, processing, and governing datasets while addressing legal, technical, and ethical challenges with case studies like EleutherAI's Common Pile.
• It advocates for future multisectoral collaboration and community-driven contributions to democratize AI research and enrich LLM training outcomes.

Best Practices for Open Datasets in LLM Training

The paper "Towards Best Practices for Open Datasets for LLM Training" reflects a comprehensive examination of the development and governance of open datasets for training LLMs. This paper emerged from a collaboration between Mozilla and EleutherAI, who convened a wide array of scholars and practitioners in 2024. The authors seek to establish normative principles and technical strategies that address the legal and ethical complexities involved in curating openly licensed datasets for LLMs.

Context and Motivation

The transparency of datasets used in training LLMs is pivotal for accountability, particularly given the opaque practices that some major companies have adopted concerning their training data. Indeed, the AI ecosystem has faced significant criticism and multiple legal actions due to suspected exploitative data practices, notably regarding copyright issues. The authors argue that the absence of openly accessible and public domain datasets effectively hinders innovation and transparency across AI research and development. By documenting effective open dataset practices, the paper aims to remedy this transparency gap and encourage a culture of openness in AI development.

Challenges in Open Dataset Development

The paper identifies substantial challenges facing the development of open datasets. It cites issues with the legal variability of copyright laws across jurisdictions, the often incomplete metadata accompanying datasets, and the substantial resource investment necessary for digitizing and processing this data. Moreover, the collaboration of legal, technical, and policy experts is necessary to navigate and integrate these datasets effectively. The document posits that several of these challenges echo early difficulties in the open-source software landscape, such as issues of data quality and the reliance on community-driven contributions.

Guiding Principles

Through the convocation, seven guiding principles were identified to aid in developing open datasets:

1. Fostering a competitive and transparent ecosystem of LLMs.
2. Enhancing accountability and transparency through dataset reproducibility.
3. Minimizing harm and incorporating preference signals during data collection.
4. Improving diversity by representing global languages and different cultural viewpoints in datasets.
5. Establishing reciprocity to ensure mutual benefits for data contributors.
6. Engaging with like-minded organizations in the domain of open data.
7. Preserving datasets for future accessibility and use.

These principles are intended to guide the community toward consistent and shared practices that can improve trustworthiness and openness in LLM training datasets.

Practical Recommendations and Case Studies

The paper provides practical recommendations for best practices in sourcing, processing, governing, and releasing open datasets. These include encoding preferences in metadata, prioritizing high-quality data sourcing, ensuring compliance with existing transparency standards, and tailoring data governance to specific use cases.

To illustrate best practices, the paper highlights the case studies of EleutherAI's "Common Pile" and Pleias' "Common Corpus" and "YouTube-Commons." These examples provide insights into practical actions taken to compile comprehensive, open-access datasets. The projects underscore the necessity of overcoming technical challenges, such as ensuring accurate optical character recognition (OCR), and the significance of community input and improvement over time.

Implications and Future Directions

The creation and maintenance of open datasets carry implications for the future of AI development, stressing the necessity of multisectoral collaboration to preserve the web's openness. These initiatives could democratize access to high-quality training data, allowing smaller entities and researchers to compete with established tech giants. Furthermore, the paper emphasizes the importance of comprehensive, inclusive datasets that accurately represent the diversity of human language and culture.

Future work must address how to foster sustainable and scalable open dataset contributions and encourage the involvement of underrepresented communities in dataset creation and curation. By doing so, the community can continue to enrich the quality and representativeness of training datasets.

The paper ultimately seeks to lay a foundation for an open LLM development ecosystem where transparency, accountability, and diversity are prioritized, ensuring that AI advancements are beneficial to a broader societal range.