Contribute

AI Attributes curates open models that read attributes of text. You can use them, build on them, and contribute a fine-tuned model of your own.

Contribute a model

A contributed model is listed beside ours on its product's Models page and offered through the API, under its author's name. These are the draft rules; they will be revised as the first contributions come in.

1. One attribute, one label set. A model reads one attribute and scores the label set its product publishes, in the same order. A new attribute, or a different label set, starts as a proposal first.
2. Open weights on Hugging Face. The weights are public in your own repository. You stay the author and keep the credit.
3. A permissive licence. MIT, Apache-2.0, BSD or similar, for the weights and for anything they were built on. No non-commercial or copyleft terms.
4. Training data you may use. Name every dataset and its licence. No private, scraped-without-permission or personal data.
5. A model card. What it was fine-tuned from and the exact revision, the data, the training settings, the test results for every label, and its known weaknesses.
6. The same test. Results are reported on the product's held-out test set, chosen before the test was run. Synthetic training data is declared.
7. Pinned. A listed model is a specific revision. A new version is a new entry, so old scores stay traceable.
8. Honest wording. No claim that the model measures what anyone feels, judges right and wrong, or shows that a system is conscious.

How a model is reviewed: we check the card and the licences, run the model on the product's test set, and compare it with the models already listed. A model that passes is listed with its results. One that doesn't gets the reasons.

Use the products

• OpenEmotions^™: Open models for emotions in text: the 28 GoEmotions labels.
• OpenSensations^™: Open models for the senses in text: vision, hearing, smell, taste, touch and the inner body.
• OpenMorals^™: Open models for moral framing in text, following Moral Foundations Theory.
• OpenSimulations^™: Simulated situations where a model acts: tools, a deadline, and every decision executed and logged.
• AI Observability Studio: Run one prompt through many models and watch them decide, side by side.
• Am I Conscious?: An open experiment observing one persistent AI agent over time.

Build on them

Each open site has an API guide and a Models page under its Documentation. Call a model through the API with a URL and a key, or download the weights from Hugging Face and run them yourself. The full list is on Models.

Help with the research

Expertise in AI, psychology, affective science, moral psychology, sensory science or philosophy of mind is welcome, and so are careful questions from anyone. The most useful notes say which text or run a result came from, what you expected, and why.

A public channel for contributions isn't open yet. This page will say where to send them when it is.

Before contributing, read the principles and the method and ethics.