Struct PretrainConfig 

pub struct PretrainConfig {

    pub dataset_path: PathBuf,
    pub tokenizer_dir: PathBuf,
    pub run_dir: PathBuf,
    pub lr_max: f32,
    pub lr_min: f32,
    pub warmup_steps: usize,
    pub total_steps: usize,
    pub batch_size: usize,
    pub seq_length: usize,
    pub steps_per_epoch: usize,
    pub seed: u64,
    pub grad_clip: f32,
    pub weight_decay: f32,
    pub target_val_loss: f32,
    pub patience_epochs: usize,
    pub min_epochs_before_early_stop: usize,
    pub regime: TrainingRegime,
}

Pretraining configuration — directly maps to CLI flags plus the convergence-policy block from the contract.

Fields

dataset_path: PathBuf

--dataset — path to tokenized shard index or raw corpus.

tokenizer_dir: PathBuf

--tokenizer — directory containing vocab.json + merges.txt.

run_dir: PathBuf

--output-dir — training run root.

lr_max: f32

Peak learning rate (after warmup).

lr_min: f32

Minimum learning rate at end of cosine decay.

warmup_steps: usize

Number of warmup steps.

total_steps: usize

Total training steps (including warmup).

batch_size: usize

Micro-batch size.

seq_length: usize

Sequence length per example.

steps_per_epoch: usize

How many steps per epoch — the driver flushes per-epoch artifacts every steps_per_epoch steps.

seed: u64

Fixed seed for INV-TRAIN-006 reproducibility.

grad_clip: f32

Gradient-clip max L2 norm (spec default 1.0).

weight_decay: f32

AdamW weight decay.

target_val_loss: f32

GATE-TRAIN-003 target final val_loss.

patience_epochs: usize

Patience for convergence / early-stop (contract default 2).

min_epochs_before_early_stop: usize

Minimum epochs before early-stop can trigger (contract default 3).

regime: TrainingRegime

INV-TRAIN-005 regime — selects the epoch-zero val_loss cap. Defaults to Finetune (10.0) to preserve v1.1.0 behavior.

Implementations

impl PretrainConfig

pub fn model_2_defaults( dataset_path: PathBuf, tokenizer_dir: PathBuf, run_dir: PathBuf, ) -> Self

Recipe that aligns with the MODEL-1 v2 post-mortem remedy: LR=5e-5, rank=32 (moot here — not LoRA), seed=42. Defaults regime to Finetune (MODEL-1-style). Use [PretrainConfig::from_scratch] to flip to the MODEL-2 regime.

Trait Implementations

impl Clone for PretrainConfig

fn clone(&self) -> PretrainConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for PretrainConfig

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

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for PretrainConfig

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

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Serialize for PretrainConfig

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

where __S: Serializer,

Serialize this value into the given Serde serializer.