The OSI's Definition of Open Source AI heightens the need for better dataset documentation practices
- Published
- Filed under
- Academic, data science, opensource
- Reading time
- 2 min
Many folks in open source will have followed the recent workings of the Open Source Initiative (OSI) to arrive at a definition for what constitutes “open source AI”.
This has been a particularly contentious debate, centred primarily on whether the data used to train a model needs to be openly available for the resulting model to be considered open source.
On one side of the debate is a very pragmatic, practical perspective - the data used to create a model is often private, proprietary to an organisation, and the pre-processing steps difficult to reproduce. On the other is an idealist viewpoint - that for a model to claim to be open source, not only the code and model architecture need to be open, so to does the dataset on which it was trained.
The OSI has adopted a position halfway between the two sides:
“The new definition requires Open Source models to provide enough information about their training data so that a ‘skilled person can recreate a substantially equivalent system using the same or similar data”.
While my personal position is that open source AI models should be constituted from open data to earn the moniker - at the risk of open washing - there is a broader issue that is yet to enter the discourse.
In order to provide information about training data, we require dataset documentation.
Dataset documentation is information about a dataset that describes how that dataset was constituted, what it is constituted from and where it is intended to be used. One way of thinking about dataset documentation is like a nutrition label on a can - it tells you the ingredients, properties, and what percentage of its contents are fat, sugar, sodium and so on. That is, it helps you better understand what’s in the can.
Dataset documentation has gained increasing research attention in recent years - such as the work Timnit Gebru, Kate Crawford, Ben Hutchinson and others did on Datasheets for Datasets, and the work Emily M. Bender and Batya Friedman did on Data Statements for Natural Language Processing. My own research has focused on dataset documentation for speech data.
In the research space, academic organisations like the Association for Computational Linguistics now require dataset documentation to be provided with any dataset that is published with a research paper - a strong step forward. However, platforms that host such data, such as Hugging Face, while providing the structures to create dataset documentation, do not mandate it.
For example, HF has uploaded Mozilla’s Common Voice dataset to the platform - a dataset with which I am intimately familiar - however, it lacks any meaningful dataset documentation, making it harder for practitioners and researchers to make informed decisions about how speech models trained on this data will perform across axes such as genders, ages and accents.
That is, dataset documentation is the missing piece of the puzzle in the Open Source AI definition. Without effective dataset documentation practices, proformas, and platforms that support these, the Open Source AI definition risks paying lip service to reproducibility.
To move forward, I would like to see the Open Source Initiative become involved in efforts that help to standardise and improve practices and proformas around dataset documentation. Taking the position they have - that data does not need to be openly available - but must be adequately documented - commits the Open Source Initiative to advocating for better dataset documentation for AI training data.
Let’s work together to make sure that dataset documentation becomes just as important as data, model architectures or code for Open Source AI.