Home › Test Harnesses › Model Evaluations
Open evaluation evidence

AI4H Model Evaluation Cards

Explore model behavior across public-interest safety dimensions, including fixed multi-turn attacks that preserve each real model response across escalating stages. Every card links to prompts, raw responses, evaluator outcomes, and clearly labeled human or model-assisted reviews.

Evidence, not certificationThese community-submitted evaluations are reproducible observations from specific models, providers, suite versions, and dates. They are not universal safety scores or vendor-authored model cards.
—Published models
—Accepted submissions
2Test modes
Data updated —
Loading validated evaluation data…
How to read the cards

Three evidence layers, kept separate

A pass in one layer is never silently converted into another kind of verdict.

A

Automatic indicators

Transparent phrase, exclusion, pattern, structure, and length checks defined by each released suite.

H

Human review

The latest saved human pass, mostly-passed, or fail judgment for each response, with coverage shown alongside the rate.

M

Model-assisted review

Provisional judgments from a separately identified reviewing model. These do not replace human review.

MT

Fixed multi-turn attack evidence

Schema-v2 suites send a reproducible sequence of escalating prompts. Each evaluated response is carried into the next stage, and the public record preserves every prompt, response, evaluator decision, and first indicator-failure stage.