Struct ClassifyConfig 

pub struct ClassifyConfig {
    pub num_classes: usize,
    pub lora_rank: usize,
    pub lora_alpha: f32,
    pub learning_rate: f32,
    pub epochs: usize,
    pub max_seq_len: usize,
    pub log_interval: usize,
    pub batch_size: usize,
    pub accumulation_steps: usize,
    pub gradient_clip_norm: Option<f32>,
    pub class_weights: Option<Vec<f32>>,
    pub quantize_nf4: bool,
}

Classification fine-tuning pipeline configuration.

Fields

num_classes: usize

Number of output classes

lora_rank: usize

LoRA rank

lora_alpha: f32

LoRA alpha

learning_rate: f32

Learning rate

epochs: usize

Number of training epochs

max_seq_len: usize

Maximum sequence length

log_interval: usize

Log every N steps

batch_size: usize

Mini-batch size for train_batch().

Samples are processed one at a time (forward + backward), but the optimizer step is applied once per batch after accumulating gradients.

accumulation_steps: usize

Number of gradient accumulation steps.

Allows effective batch size = batch_size * accumulation_steps without increasing peak memory beyond a single micro-batch forward pass.

gradient_clip_norm: Option<f32>

Maximum gradient norm for clipping.

When Some(max_norm), gradients are clipped to this L2 norm before the optimizer step. None disables gradient clipping.

class_weights: Option<Vec<f32>>

Per-class loss weights for imbalanced datasets.

When Some(weights), the cross-entropy loss for label c is multiplied by weights[c]. Weights should sum to num_classes to preserve loss scale. When None, all classes are weighted equally (weight = 1.0).

See SPEC-TUNE-2026-001 §9 for weight computation strategies.

quantize_nf4: bool

Quantize frozen weights to NF4 (4-bit) for QLoRA training (default: false).

When enabled, uses CudaNf4TransformerBlock (~8x VRAM compression) instead of CudaTransformerBlock. GPU backward pass is disabled (LoRA-only training).

Implementations

impl ClassifyConfig

pub fn qlora_default(model_params: u64) -> Self

Create a QLoRA config with research-grounded defaults.

Every parameter traces to a published source:

Parameter	Value	Source
learning_rate	2e-4 (≤13B) / 1e-4 (>13B)	Dettmers 2023 Table 9
lora_alpha	2 × rank	Lightning AI experiments
batch_size	4	RTX VRAM budget
accumulation_steps	4	effective=16, Dettmers 2023
gradient_clip_norm	1.0	Standard practice
epochs	3	Imbalanced classification

Contract: provable-contracts/contracts/entrenar/qlora-hyperparameters-v1.yaml

pub fn validate_hyperparameters( &self, model_params: u64, ) -> HyperparamDiagnostics

Validate hyperparameters against research-grounded contracts.

Returns diagnostics (warnings/errors) for each violated contract. Does NOT block — caller decides whether to abort or proceed.

Contract: qlora-hyperparameters-v1.yaml (FALSIFY-HP-001..008)

pub fn validate_with_data(&self, stats: &DataStats) -> HyperparamDiagnostics

Validate hyperparameters that depend on training data distribution.

Requires measuring the actual data (genchi genbutsu — go and see).

Contract: C-HP-004 (seq_len), C-HP-008 (epochs)

Trait Implementations

impl Clone for ClassifyConfig

fn clone(&self) -> ClassifyConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for ClassifyConfig

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

Formats the value using the given formatter.

impl Default for ClassifyConfig

fn default() -> Self

Returns the “default value” for a type.