1. Core Training Parameters

per_device_train_batch_size

• What it does:

  Number of samples processed at once on each GPU/device.

• Why it matters:

  Larger batches use more GPU memory but train faster. Smaller batches save memory but may be less stable.

• Example:

  per_device_train_batch_size=2 → Each GPU processes 2 samples per forward/backward pass.

gradient_accumulation_steps

• What it does:

  How many batches to process before updating the model weights (sums gradients across steps)

• Why it matters:

  Mimics a larger batch size without crashing your GPU.

• Example:

  gradient_accumulation_steps=4 with per_device_train_batch_size=2 → Effective batch size = 2 × 4 = 8.

num_train_epochs

• What it does:

  Total number of times the model sees the entire training dataset.

• Why it matters:

  Too few → model underfits. Too many → overfits.

• Example:

  num_train_epochs=3 → Train for 3 full passes through the data.

learning_rate

• What it does:

  Step size for updating weights.

• Why it matters:

  Too high → unstable training. Too low → slow progress.

• Example:

  learning_rate=2e-5 → Common starting point for fine-tuning LLMs.

max_grad_norm

• What it does:

  Clips gradients to prevent them from exploding (sudden large updates).

• Why it matters:

  Stabilizes training, especially for deep networks.