Fine-tuning and Efficient Adaptation

Covers transfer learning paradigm, pre-training/fine-tuning split, what transfers, transfer learning efficiency.

Covers full fine-tuning procedure, fine-tuning hyperparameters, learning rate selection, batch size effects.

Covers forgetting phenomenon, forgetting measurement, forgetting mitigation, pre-trained capability preservation.

Covers discriminative fine-tuning, layer-wise learning rates, warmup for fine-tuning, learning rate decay.

Covers few-shot fine-tuning, data augmentation, sample efficiency patterns, small data strategies.

Covers parameter storage costs, multi-task deployment, PEFT efficiency, PEFT quality trade-offs.

Covers weight update decomposition, low-rank assumption, LoRA efficiency gains, LoRA flexibility.

Covers LoRA formulation W + BA, rank selection, initialization scheme, LoRA gradient computation.

Covers LoRA module design, merging weights, LoRA training loop, LoRA in PyTorch, HuggingFace PEFT usage.

Covers rank selection guidelines, alpha/rank ratio, which layers to adapt, LoRA dropout.

Covers 4-bit quantization for base model, NF4 data type, double quantization, QLoRA memory savings.

Covers importance-based pruning, SVD-based adaptation, dynamic rank, AdaLoRA training procedure.

Covers IA3 formulation, learned rescaling vectors, IA3 parameter efficiency, IA3 vs LoRA.

Covers prefix tuning formulation, prefix length selection, prefix tuning for generation, prefix vs LoRA.

Covers prompt tuning formulation, prompt initialization, prompt tuning scaling, prompt length effects.

Covers adapter architecture, adapter placement, adapter dimensionality, adapter fusion.

Covers performance comparison, parameter efficiency comparison, task suitability, practical recommendations.

Covers instruction tuning motivation, instruction format design, instruction diversity, instruction quality.

Covers human annotation, template-based generation, seed task expansion, quality filtering.

Covers self-instruct procedure, instruction generation, response generation, filtering strategies.

Covers prompt templates, system messages, multi-turn format, chat templates, role definitions.

Covers instruction tuning data mixing, training hyperparameters, loss masking, multi-task learning.

Covers instruction following benchmarks, human evaluation, automatic evaluation, instruction difficulty.