Struct TrainingOutput 

pub struct TrainingOutput {

    pub convergence_records: Vec<IterationRecord>,
    pub final_lower_bound: f64,
    pub final_upper_bound: Option<f64>,
    pub final_gap_percent: Option<f64>,
    pub final_upper_bound_std: Option<f64>,
    pub final_upper_bound_kind: String,
    pub iterations_completed: u32,
    pub converged: bool,
    pub termination_reason: String,
    pub total_time_ms: u64,
    pub cut_stats: RowPoolStatistics,
    pub cut_selection_records: Vec<RowSelectionRecord>,
    pub worker_timing_records: Vec<WorkerTimingRecord>,
    pub training_solve_stats: MetadataTrainingSolveStats,
}

Aggregate type carrying all training data needed for output writing.

Constructed by the solver after training completes and passed to write_results. All convergence records and summary statistics are held here so the writer can read them without contacting the solver.

Fields

convergence_records: Vec<IterationRecord>

Ordered convergence records — one entry per completed iteration.

final_lower_bound: f64

Lower bound value reported after the final iteration.

final_upper_bound: Option<f64>

Upper bound value reported after the final iteration, if available.

None when no upper-bound evaluation was performed.

final_gap_percent: Option<f64>

Relative gap between final upper and lower bounds as a percentage.

None when the lower bound is zero/negative or final_upper_bound is None.

final_upper_bound_std: Option<f64>

Standard deviation of the final upper-bound estimate, if available.

None when no upper-bound evaluation was performed or the bound is exact. The value is carried separately in final_upper_bound.

final_upper_bound_kind: String

Upper-bound regime for the whole run: "statistical" (sampled forward) or "exact" (enumerated forward). Mirrored into training/convergence.parquet and training/metadata.json.

iterations_completed: u32

Number of iterations completed before the stopping condition was triggered.

converged: bool

true when training converged within the configured tolerance.

termination_reason: String

Human-readable description of the rule that terminated training.

total_time_ms: u64

Total elapsed wall-clock time for the entire training run (ms).

cut_stats: RowPoolStatistics

Summary row pool statistics for the run.

cut_selection_records: Vec<RowSelectionRecord>

Per-stage row-selection records for Parquet output.

Empty when row selection is disabled. When non-empty, written to training/cut_selection/iterations.parquet.

worker_timing_records: Vec<WorkerTimingRecord>

Per-worker timing records for training/timing/iterations.parquet.

Each entry is either a rank-aggregated row (worker_id = None) or a per-worker row (worker_id = Some(w)). Empty when timing data was not collected (e.g. single-threaded runs without the instrumentation wired). Written in iteration-major order: rank-aggregated row first, then per-worker rows sorted by (rank, worker_id).

training_solve_stats: MetadataTrainingSolveStats

Aggregate solve statistics for the training run.

Default-constructed (all fields None) by producers that do not yet record solve statistics; populated downstream and persisted into training/metadata.json by the metadata writer.

Trait Implementations

impl Clone for TrainingOutput

fn clone(&self) -> TrainingOutput

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for TrainingOutput

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.