Wikimedia's machine learning team supports models that help editors, while leaving every Wikipedia edit to human community members.
2
Chris Albon wants researchers and engineers to deploy even baseline models early, so the full production workflow is tested before the model is improved.
3
Wikimedia requires open-source infrastructure where possible, which leads the team to use self-hosted systems and AMD GPUs despite weaker tooling support.
Summary
Chris Albon describes how Wikimedia's six-person machine learning team supports models used across the Foundation. These models predict article quality, flag potentially damaging edits, and suggest links to new editors on mobile devices. The team does not edit Wikipedia. It gives editors information and suggestions, while the communities decide what changes to make. Albon wants researchers and engineers involved together from the start, with even simple baseline models deployed early through a complete workflow. Wikimedia's open-source policy shapes the infrastructure, from self-hosted servers and Gerrit to Kubernetes and Kubeflow. It also creates practical problems, especially around GPU support and the lack of mature open-source alternatives to CUDA. Albon discusses model governance through community discussion pages and model cards. He is candid that open source alone does not make models accessible when the code and infrastructure are hard to reproduce.
Wikimedia's models assist editors without making edits
Chris Albon says the Foundation hosts models that mostly help editors. They predict article quality from community-trained data and classify edits as productive, damaging, spam, or possible vandalism. The models direct editor attention rather than changing Wikipedia themselves. He describes Add-a-Link as a mobile-first example. The system suggests that a word in one article should link to another article, and a new editor can accept or reject the suggestion from a phone. An accepted suggestion becomes the editor's change, which other editors can discuss or revert. Albon sees this as a way for people to contribute from mobile devices while keeping control with the community.
Researchers and engineers should deploy a complete workflow from the start
Albon wants his engineering team involved when a product team first proposes an ML feature, alongside researchers, designers, translators, and community relations staff. He argues that a model should reach production as soon as there is even a baseline that returns true. The point is to test the full path through deployment, a public API, and documentation before spending months improving the model. He compares this to agile product development: start with one complete, imperfect workflow, then improve its accuracy, speed, documentation, and scale. This avoids handing engineers a Jupiter notebook after research is finished and asking them to reconstruct the production system.
Policies enforced in infrastructure reduce dependence on individual engineers
Neal Lathia and Chris Albon discuss the risk that only one engineer knows how a system works. Albon's preferred answer is to enforce deployment requirements in the infrastructure. A model must pass defined tests and follow a particular structure, with no skipped steps or special cases. That makes the process less dependent on an individual's habits. He admits this is imperfect. Supporting an unusual library with a custom Kubeflow service may solve an immediate problem while leaving the next engineer with unfamiliar code. The team must balance close embedding with product teams against keeping ML knowledge shared across a small central group.
Wikimedia's open-source policy shapes every infrastructure decision
Albon says Wikimedia uses open source as the default for infrastructure because the Foundation's work is funded by donors and its mission is based on open knowledge. The team uses bare-metal servers, runs its own data centers, and installs the stack rather than relying on AWS. It uses Gerrit instead of GitHub and is moving toward self-hosted GitLab. The policy affects hardware as well. AMD GPUs have better open-source support than NVIDIA GPUs, but the documentation for connecting AMD GPUs to Kubeflow is much thinner. This creates extra work for a six-person team, which also has to maintain the specialized knowledge of Wikimedia's self-hosted systems.
ML infrastructure sits between data engineering and model serving
Wikimedia keeps sensitive data inside an analytics VLAN that is not accessible from the public internet. Albon describes an architecture with one Kubernetes cluster inside that network for training and another outside it for serving, with a controlled path for moving models. Pre-processing complicates the boundary because an inference request may need data or cached features from inside the protected network. The team plans to build shared infrastructure around model registries, prediction caches, and offline and online feature stores. Albon says this belongs between ML and data engineering, especially because several members of his team came from data engineering. He also argues that serving a model involves versioning, training data, evaluation data, deployment charts, staging, and development systems.
Kubeflow provides testing capabilities that make iteration faster
Albon says Kubeflow was a major reason Wikimedia chose it as the next open-source ML infrastructure. The previous system did not support the same experimentation workflow, while Kubeflow offered capabilities such as shadow deployments and A/B testing. Shadow testing lets the team collect results from a new model without changing production behavior. Albon believes repeated testing makes difficult work manageable because teams improve through faster iterations. He also values the Kubeflow contributor community, since Wikimedia's small team has received help from contributors after discussing problems in livestreams and team conversations.
Community governance determines whether a model should run
Albon is especially proud of models trained by Wikimedia communities for their own language editions. A French Wikipedia community can define what article quality means for French Wikipedia, and other communities may have different standards. Add-a-Link demonstrates why local governance matters: German Wikipedia asked Wikimedia to turn the feature off because its linking policy differs from Arabic Wikipedia's policy. Albon wants every hosted model to have a wiki page containing its purpose, training score, code, training data, intended use, and limitations. The associated talk page would let the affected community discuss the model, raise concerns, and decide whether it should remain active. Technical safeguards alone cannot answer questions about community consent or governance.
Open source becomes useful only when people can reproduce and change the models
Albon distinguishes between code being publicly available and being accessible. Wikimedia's older model infrastructure required contributors to run the whole system locally, and he says interns took weeks to reproduce existing models. His goal is to move models into GitLab in a simple form, such as a visible scikit-learn project, so an undergraduate student or community contributor can inspect and improve it. This also applies to research. Research code may solve an interesting question but still be hard to operate with reliable uptime, efficient execution, and public data handling. Albon sees the models Wikimedia hosts as open-source projects, with accessibility measured by whether people can understand them and contribute to them.
"Open source is a means to create accessibility and where we are now with the models is that things are open source but because they're very difficult to replicate they are not accessible."Chris Albon1:00:35
Who should watch
You run a small ML platform team and need to reduce handoffs between research, engineering, and product groups.
Your organization is considering self-hosted or open-source infrastructure and wants to understand the operational costs, especially around hardware and staff knowledge.
You need a practical example of community governance for models that affect different language or user communities.
