Model Training

Split data into training, validation, and test sets

Train ML model with hyperparameter tuning

Grid search, random search, or Bayesian optimization

K-fold, stratified, or time-series cross-validation

Update all model parameters on task-specific data

Low-Rank Adaptation - add small trainable matrices to attention layers

Quantized LoRA - 4-bit quantization + LoRA for memory efficiency

Insert small trainable modules between frozen transformer layers

Learn continuous soft prompts prepended to each layer

Learn task-specific prompt embeddings (input layer only)

Infused Adapter - learn rescaling vectors for activations

Fine-tune on instruction-following datasets (e.g., Alpaca, ShareGPT)

Reinforcement Learning from Human Feedback with reward model

Train reward model from human preference comparisons

Direct Preference Optimization - simplified RLHF without reward model

Odds Ratio Preference Optimization - single-stage SFT + preference

Self-improvement via AI feedback based on constitutional principles

Freeze base model, train only the classification head

Adapt pretrained model to a new domain (e.g., medical, legal)

Further pretrain on domain-specific corpus before fine-tuning

Train smaller student model to mimic larger teacher model

Train on multiple tasks simultaneously with shared representations

Reuse pretrained model knowledge for new tasks via fine-tuning or feature extraction

Iteratively select most informative samples for labeling to minimize annotation cost

Reduce model precision (FP32→INT8/INT4) for faster inference and smaller footprint