Struct TrainingParams 

pub struct TrainingParams {

    pub epochs: usize,
    pub grad_clip: Option<f32>,
    pub lr_scheduler: Option<String>,
    pub warmup_steps: usize,
    pub save_interval: usize,
    pub output_dir: PathBuf,
    pub mode: TrainingMode,
    pub gradient_accumulation: Option<usize>,
    pub checkpoints: Option<usize>,
    pub mixed_precision: Option<String>,
    pub scheduler_params: Option<HashMap<String, Value>>,
    pub max_steps: Option<usize>,
    pub seed: Option<u64>,
    pub max_checkpoints: usize,
    pub shuffle: bool,
    pub curriculum: Option<Vec<CurriculumStage>>,
    pub profile_interval: usize,
    pub deterministic: bool,
    pub eval_interval: usize,
    pub patience: usize,
    pub distributed: Option<DistributedSpec>,
}

Training hyperparameters

Fields

epochs: usize

Number of epochs

grad_clip: Option<f32>

Gradient clipping threshold

lr_scheduler: Option<String>

Learning rate scheduler

warmup_steps: usize

Warmup steps

save_interval: usize

Save checkpoint every N epochs

output_dir: PathBuf

Output directory for checkpoints

mode: TrainingMode

Training mode (regression or causal_lm) ENT-114: Uses CausalLMLoss when mode=causal_lm

gradient_accumulation: Option<usize>

Gradient accumulation steps (for large batch simulation)

checkpoints: Option<usize>

Number of gradient checkpoints (for memory optimization)

mixed_precision: Option<String>

Use mixed precision training (bf16 or fp16)

scheduler_params: Option<HashMap<String, Value>>

Scheduler-specific parameters (t_max, gamma, step_size, etc.)

max_steps: Option<usize>

Maximum training steps (overrides epochs if set)

seed: Option<u64>

Global random seed for reproducibility

max_checkpoints: usize

Maximum number of checkpoints to keep (default 5, 0 = unlimited)

shuffle: bool

Shuffle training batches per epoch (default true)

curriculum: Option<Vec<CurriculumStage>>

Multi-stage data mixing: switch data sources at step boundaries. Each stage specifies a data path and the step at which to transition.

profile_interval: usize

Print per-phase timing breakdown every N steps (0 = disabled)

deterministic: bool

Enable bitwise deterministic training mode (C-DETERM-001). Sets CUBLAS_WORKSPACE_CONFIG, cuDNN deterministic mode, disables cuDNN benchmark. May reduce throughput but guarantees reproducibility.

eval_interval: usize

Steps between validation evaluations (defaults to save_interval if 0). Decouples eval frequency from checkpoint save frequency.

patience: usize

Number of eval intervals without improvement before early stop (0 = disabled).

distributed: Option<DistributedSpec>

Distributed data-parallel training configuration. When present, enables multi-GPU / multi-node DDP pretraining.

Trait Implementations

impl Clone for TrainingParams

fn clone(&self) -> TrainingParams

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for TrainingParams

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

Formats the value using the given formatter.

impl Default for TrainingParams

fn default() -> Self

Returns the “default value” for a type.

impl<'de> Deserialize<'de> for TrainingParams

where TrainingParams: Default,

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Serialize for TrainingParams

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.