distill::convergence::ConvergenceState¶

Public Attributes¶

	Name
List	losses
float	best_loss
int	steps_since_improvement
bool	converged
bool	hard_stopped
bool	warning_issued

Detailed Description¶

class distill::convergence::ConvergenceState;

Mutable tracking state for a single training run.

Attributes:
    losses: Full loss history for the current run.
    best_loss: Best (lowest) loss observed so far.
    steps_since_improvement: Consecutive steps without a min_delta improvement.
    converged: True when loss is below target and patience is exhausted.
    hard_stopped: True when loss exceeded the hard-stop threshold.
    warning_issued: True when the warning threshold has been breached at least once.

Public Attributes Documentation¶

variable losses¶

static List losses =  field(default_factory=list);

variable best_loss¶

static float best_loss =  float("inf");

variable steps_since_improvement¶

static int steps_since_improvement =  0;

variable converged¶

static bool converged =  False;

variable hard_stopped¶

static bool hard_stopped =  False;

variable warning_issued¶

static bool warning_issued =  False;

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