← Reference notes

License: CC BY 2.0 Belgium. Source: https://archive.fosdem.org/2025/schedule/event/fosdem-2025-6020-community-insights-best-practices-for-open-datasets-for-llm-training/

Summary

Kasia Odrozek (Mozilla), speaking at FOSDEM on 2 February 2025, describes why large labs stopped publishing the sets they train on after copyright suits, and what a more open practice would look like. She sorts openness into three levels: documentation enough for someone else to try to repeat the work, a download with no license claim, and a set people may reuse, share, and change, including its pieces. Building that third kind from openly licensed and public-domain material is hard. License statements on crawled pages are incomplete, public-domain copies are often not digitized, PDF text is messy to extract, and a small group may have no legal entity behind the release. A consent fight is also blocking researchers, including through robots.txt. The recommendations she walks through include preference metadata, written documentation, toxicity classifiers, data trusts, opt-out interfaces, a safer harbor for small labs, machine-readable license and consent fields, and public-goods funding. Attribution of the Creative Commons items inside a set is still unsolved; one attempt is a column in the data file. This note follows the talk captions. It does not use the research paper linked from the schedule page.

Why it is here

The library is asked how to keep personal data safe, and the assistant notes already assume models were trained on somebody's writing. This talk is a public account of what "open" training data does and does not mean, and why consent and license scope still matter after a file is posted.

@misc{iil_note_3ke71bqk,
  author = {{iinventedlight Library}},
  title = {{Community Insights: Best Practices for Open Datasets for LLM Training}},
  howpublished = {iinventedlight Library},
  year = {2026},
  date = {2026-10-18},
  url = {https://iinventedlight.com/library/reference/talks/fosdem-2025-open-datasets-llm-training/},
  note = {Permanent ID iil:note:3ke71bqk. License: CC-BY-2.0. Summary of https://archive.fosdem.org/2025/schedule/event/fosdem-2025-6020-community-insights-best-practices-for-open-datasets-for-llm-training/.},
}

For language models and agents: start at /llms.txt. This page in Markdown: /library/reference/talks/fosdem-2025-open-datasets-llm-training.md.