Principle:Googleapis Python genai Tuning Job Configuration
| Knowledge Sources | |
|---|---|
| Domains | Fine_Tuning, Hyperparameter_Tuning |
| Last Updated | 2026-02-15 00:00 GMT |
Overview
A parameter set that controls the hyperparameters, method, and naming of a model fine-tuning job.
Description
Tuning Job Configuration specifies how a fine-tuning job should be executed. This includes the training method (supervised fine-tuning, distillation, or preference tuning), hyperparameters (epoch count, learning rate multiplier), LoRA adapter size, output model naming, and optional validation dataset. These parameters directly influence training duration, model quality, and resource consumption. Proper configuration balances model quality against cost and training time.
Usage
Configure tuning jobs based on dataset size, quality requirements, and budget. For small datasets (hundreds of examples), use fewer epochs to avoid overfitting. For larger datasets, increase epochs. Use adapter_size to control the LoRA rank (smaller = more efficient, larger = more expressive). Set a descriptive tuned_model_display_name for model management.
Theoretical Basis
Fine-tuning hyperparameters control the optimization process:
- Epoch count: Number of complete passes over the training data. More epochs = more learning but risk of overfitting
- Learning rate multiplier: Scales the base learning rate. Higher values = faster learning but risk of divergence
- Adapter size (LoRA): Controls the rank of low-rank adaptation matrices
For LoRA fine-tuning, instead of updating all model weights W, a low-rank update is applied:
Where B and A are low-rank matrices with rank r (the adapter size). Smaller r means fewer trainable parameters.