Alignment
Model Spec
2024ActiveUpdated: 15 September 2026Published
Key
innovation
An explicit document specifying a model's desired behavior that becomes the source of truth for tuning, alignment and evaluation.
Category
Alignment
Abstraction level
Pattern
Operation level
Post-trainingApplication
Use cases
Model alignmentDefining model behavior policiesModel behavioral evaluation and testing
How it works
A team defines rules and priorities (e.g. safety over helpfulness in specified situations) along with examples and exceptions. The spec is then used to generate training data, tune the model, build behavioral tests and assess whether the model meets the intent. Updates to the spec feed into subsequent alignment iterations.
Problem solved
Model behavior is often implicit and inconsistent. A Model Spec makes the expected behavior explicit, measurable, enforceable and auditable.
Evolution
Original paper · 2024
OpenAI Model Spec
2024
OpenAI publishes its Model Spec
Inflection pointThe first widely known explicit specification of a model's desired behavior.