diff --git a/README.md b/README.md index 031a447..a67626c 100644 --- a/README.md +++ b/README.md @@ -4,7 +4,7 @@ Welcome to the ROOST Model Community (RMC)! The RMC’s mission is to make open The RMC seeks to add tangible value for both active and potential users of open safety models and the developers of those models (called RMC Partners). Participants in the RMC benefit from: -- **A Community of Trust**: We only partner with high quality open safety models that you can trust; see our [eligibility criteria] for model selection. The RMC also stewards community conversations to encourage continued engagement and collective problem solving, via both routine events and casual conversations in our [Discord server]. +- **A Community of Trust**: We only partner with high quality open safety models that you can trust; see our [eligibility criteria](eligibility-criteria.md) for model selection. The RMC also stewards community conversations to encourage continued engagement and collective problem solving, via both routine events and casual conversations in our [Discord server]. - **Education and Implementation Support**: We share resources to help you better understand the open safety landscape, including introductory guides to AI for Trust & Safety and evaluation outcomes. Once you’re ready to use a specific model, the RMC also provides implementation tips and documentation, such as sample datasets or direct integrations with other ROOST projects. @@ -59,7 +59,7 @@ For our purposes, we see “open safety models” as an AI model (including but Additionally, we prefer models that do not have non-commercial licenses and licenses that do not place restrictions on how outputs are handled. This is because many use cases for these models happen in commercial environments and have reporting requirements (such as CSAM). These license restrictions are incompatible with the Trust & Safety users we aim to support. -Although there are many open safety models, we hold a specific bar for RMC Partners. To formally become an RMC partner, a model must meet our [eligibility criteria]. We work closely with RMC partners to ensure that their models meet these criteria before adding them to the community. For any questions about these criteria, please feel free to reach out – we designed these criteria to meet the community’s needs and are open to feedback! +Although there are many open safety models, we hold a specific bar for RMC Partners. To formally become an RMC partner, a model must meet our [eligibility criteria](eligibility-criteria.md). We work closely with RMC partners to ensure that their models meet these criteria before adding them to the community. For any questions about these criteria, please feel free to reach out – we designed these criteria to meet the community’s needs and are open to feedback! ## RMC Partners @@ -88,5 +88,5 @@ Traditional open source development works well for software, but AI models prese Our community brings the open access and community spirit of open source software development to AI models, made possible by partnership with AI researchers and model creators. These partners commit to developing and openly releasing the weights of safety-finetuned models that are free to access, platform-agnostic in deployment, and have no commercial or output restrictions in their licenses. They also actively participate in the community, gathering feedback from practitioners and supporting their implementation journeys. In turn, ROOST cultivates a vibrant community-of-practice where safety teams share knowledge and strategies for successfully deploying these models in real-world scenarios. [discord server]: https://discord.gg/UXmBqy7kFX -[eligibility criteria]: https://docs.google.com/spreadsheets/d/1gkRwjCYFlYrah1WBZu96iWzAVUj555WO4rxR5NswzEI/edit?gid=667042538#gid=667042538 +[eligibility criteria]: eligibility-criteria.md [open an issue]: https://github.com/roostorg/model-community/issues diff --git a/eligibility-criteria.md b/eligibility-criteria.md new file mode 100644 index 0000000..49d9773 --- /dev/null +++ b/eligibility-criteria.md @@ -0,0 +1,12 @@ +# RMC Model Eligibility Criteria + +If you are unsure whether your model is eligible, please reach out to anne@roost.tools to discuss! + +### MODEL NAME: _____________ + +| Criteria | Sub-Criteria | +| --- | --- | +| **Freely Usable** | Is your model released under a permissive open license compatible with real-world safety use cases (e.g., Apache 2.0, MIT, or equivalent)? Some factors that we value include:
• The model has no monetary cost to access (e.g., downloadable weights are freely available)
• Deployment of the model is platform agnostic
• The model is allowed to be used in commercial environments
• The model does not restrict creating derivatives
• The model does not restrict sharing of outputs or artifacts with third parties, such as reporting bodies

The reason why we particularly value conventionally permissive open source licenses is that we aim to create resilient tools that cannot be taken away from their users. We take guidance from the open source community's stance on ethical use licenses such as OpenRAIL, which note that restrictive licenses open the door to that possibility and thus are both less valuable for our users and misaligned with the goals of open source. However, licenses from common open weight models such as Gemma or Llama are generally acceptable, as long as they are not overly restrictive.

Is your model packaged and documented for practical reuse (straightforward access, clear instructions, reasonable runtime requirements)? | +| **In Demand** | Is the model meaningfully differentiated (e.g., by use case, capability, etc.)? This could include the model being first of its kind, having exceptional reach, or otherwise adding value to RMC users.

Is the model's use case broadly applicable to safety workflows? | +| **Active** | Will the model be continuously updated and engaged with? This can be signalled or committed to via a clear plan for active work, such as named maintainers, target windows for upcoming updates, commitments to regular participation with the RMC community (e.g., discussions via GitHub or Discord; participation in RMC office hours), etc. | +| **Transparent** | Is the model accompanied by a model card or similar transparency artifacts? If not, will it be?

We recognize that even organizations that highly value transparency may not be able to commit to sharing each of these elements, and so we do not expect that all RMC models have perfectly comprehensive artifacts. As a result, we simply want to see best efforts towards transparency, however that may look for your organization.

For extra guidance, we look for some of the following factors to assess transparency:
1. Sample datasets or data sources
2. Policy and taxonomy definitions
3. Annotation methodology
4. Performance benchmarks
5. Bias documentation, known limits
6. Implementation guidance
7. Model license | \ No newline at end of file