distill::convergence::ConvergenceTracker¶

Public Functions¶

	Name
	init(self self, ConvergenceConfig config)
ConvergenceState	state(self self)
str	step(self self, float loss)

Protected Attributes¶

	Name
	_config
	_state

Detailed Description¶

class distill::convergence::ConvergenceTracker;

Tracks training loss, detects plateaus, and enforces two-tier stopping.

Call ``step(loss)`` after each training iteration. The return value is one of:

* ``"continue"`` — normal operation, keep training.
* ``"warning"`` — loss exceeded the warning threshold (issued once per run).
* ``"hard_stop"`` — loss exceeded the hard-stop threshold; halt immediately.
* ``"converged"`` — loss sustained below target for patience steps with no
  min_delta improvement; training is complete.

Public Functions Documentation¶

function init¶

__init__(
    self self,
    ConvergenceConfig config
)

Initialise the tracker with per-specialist convergence parameters.

Args:
    config: Convergence thresholds and plateau detection settings.

function state¶

ConvergenceState state(
    self self
)

Return the mutable tracking state (read-only for callers).```


### function step

```python
str step(
    self self,
    float loss
)

Record a loss value and return the training status.

Args:
    loss: The current distillation loss value.

Returns:
    One of ``"continue"``, ``"warning"``, ``"hard_stop"``, or ``"converged"``.

Protected Attributes Documentation¶

variable _config¶

_config;

variable _state¶

_state;

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