distill::convergence::ConvergenceConfig¶

Public Attributes¶

	Name
float	distill_loss_target
float	distill_loss_warning
float	distill_loss_hard_stop
int	patience
float	min_delta

Detailed Description¶

class distill::convergence::ConvergenceConfig;

Per-specialist convergence thresholds and plateau detection parameters.

Attributes:
    distill_loss_target: Loss must drop to this value for convergence.
    distill_loss_warning: Loss above this logs a warning and continues.
    distill_loss_hard_stop: Loss at or above this halts training immediately.
    patience: Consecutive steps without improvement before early stopping.
    min_delta: Minimum loss improvement required to reset the patience counter.

Public Attributes Documentation¶

variable distill_loss_target¶

static float distill_loss_target =  2.5;

variable distill_loss_warning¶

static float distill_loss_warning =  3.0;

variable distill_loss_hard_stop¶

static float distill_loss_hard_stop =  5.0;

variable patience¶

static int patience =  100;

variable min_delta¶

static float min_delta =  0.01;

Updated on 2026-10-09 at 03:59:43 +0000