LoRA Rank

Controls adapter capacity. High rank (64+) captures voice identity strongly but can override the model's ability to follow instruct/style prompts, producing flat delivery. Low rank (8-16) preserves expressiveness and instruct-following but captures less voice detail. Start with 16-32 for a balance.

LoRA Alpha

Scaling factor for LoRA weights. Effective strength = alpha / rank. Higher alpha relative to rank means the adapter has more influence. A common starting point is alpha = 2x rank.

Epochs

Number of full passes over the dataset. Too few: voice identity won't converge. Too many: overfitting, stuttering, repetition artifacts. 15-30 epochs is a good range for 20+ samples.

Learning Rate

Step size for weight updates. Default 5e-6 is conservative. Higher (1e-5) trains faster but risks instability. Lower (1e-6) is safer for longer training runs.

Batch Size

Samples processed per step. Limited by VRAM. Batch 1 with gradient accumulation is typical for 24GB cards.

Grad Accum Steps

Simulates larger batches by accumulating gradients over N steps before updating. Effective batch = batch_size x grad_accum. Higher values give smoother training at the cost of speed.

Training data tip: For expressive voices, include samples with varied emotions (happy, sad, angry, calm). Neutral-only training data will produce a flat voice that resists instruct prompting.