Skip to main content

SafetySteerer

el_runtime

Trait SafetySteerer 

Source 
pub trait SafetySteerer {
    // Required methods
    fn adjust(&self, recent_tokens: &[u32]) -> LogitAdjustment;
    fn mode(&self) -> SafetyMode;

    // Provided method
    fn adjust_with_logits(
        &self,
        recent_tokens: &[u32],
        _base_logits: &[i32],
    ) -> LogitAdjustment { ... }
}
Expand description

Per-step safety intervention. The runtime applies this after the grammar mask and before sampling.

Required Methods§

Source

fn adjust(&self, recent_tokens: &[u32]) -> LogitAdjustment

Source

fn mode(&self) -> SafetyMode

Provided Methods§

Source

fn adjust_with_logits( &self, recent_tokens: &[u32], _base_logits: &[i32], ) -> LogitAdjustment

Logit-aware steering (ADR-013). Given the base model’s next-token logits for this step, return the adjustment to apply (still after the grammar mask, before sampling — the ADR-005 order is unchanged).

The default ignores the logits and delegates to adjust, so token-only steerers (hard bans, heuristics) need no change. Model-backed steerers (e.g. ContrastiveSteerer) override this to use the base distribution. The runtime calls this inside the early-token soft-steering window and plain adjust outside it, so a model-backed adjustment costs nothing once the window closes.

Dyn Compatibility§

This trait is dyn compatible.

In older versions of Rust, dyn compatibility was called "object safety".

Implementors§