Skip to main content

Crate rlox_rl_ops

Crate rlox_rl_ops 

Source
Expand description

Estimator-agnostic advantage and token-KL ops for rlox.

This crate has no dependency on rlox-core (no environment physics, replay buffers, or GAE). It is intentionally slim so that rlox-sandbox can depend on it without pulling in the entire training data-plane.

§Public API

The f64 KL functions are re-exported at crate root via kl::*.

Re-exports§

pub use error::RlOpsError;
pub use estimator::AdvantageEstimator;
pub use grpo::GroupRelativeEstimator;
pub use kl::*;

Modules§

error
estimator
grpo
Group-Relative Policy Optimisation (GRPO) advantage estimator.
kl
Token-level KL divergence operations (exact and Schulman 2020 estimator).