Potluck
Contribute
Alpha is a potluck: every model, dataset and tool in it was brought by someone. The places it’s still weak — a language, a benchmark, a missing piece of the stack — are where your help counts most.
Help improve data for your language
Pick your language, then choose one way to help.
Data goes to Common Voice and the Mozilla Data Collective, in a new tab. What you add joins open releases, and can’t be pulled from copies already downloaded.
Models are weaker in some languages for several reasons: how they were trained, how they were tested, and how much text in the language they had to learn from. Data is one of those reasons, and it’s one you can help with. On Common Voice, anyone can read sentences aloud or write new ones to build an open dataset in their language. What you add goes into open releases: you can ask to be left out of future ones, but not out of copies already downloaded or used for training. If your community already has data, the Mozilla Data Collective lets it share that data on its own terms: who can use it, and for what.
Ways to help
-
Help evaluate models in your language
Benchmarks show whether a model understands a language, not whether its answers read naturally, get everyday tasks right or are safe. People who speak it can tell. Rate answers from Alpha’s models, without seeing which model wrote each one.
-
Tell us what’s missing
A language Alpha stumbles in, a tool you expected, a model you’d like to see — tell us, and it goes to the people building Alpha.
Open gaps on the roadmap
Pieces of an open source AI stack still missing or under way. Each needs partners.
- Local benchmarks Being built
Evaluation suites designed for local community use-cases.
- Benchmark evaluation datasets Being built
Independent, reproducible benchmarks and evaluation suites for foundation models, agentic systems, and safety.
- Pretraining datasets Not started
Large repositories of data for pre-training.
- Query router Not started
Intelligent routing of user queries to the most appropriate model, service, or logic.
- Model router Being built
An inference module that routes inference requests to different models based on parameters like use-case, intent, subscription class, and geography.
- Agent service Being built
A hosted service for deploying and configuring agents.