LoRA i quiz dotyczący dostrajania efektywnego pod kątem parametrów

Odpowiedzi i wyjaśnienia pojawiają się w miarę rozwiązywania quizu.

1. In the original LoRA formulation for a selected weight matrix, what is learned while that matrix stays frozen?
   • A replacement vocabulary with no weight update
   • Every entry of the full base matrix from random initialization
   • Two factor matrices forming an additive low-rank update
   • A permanent deletion of the entire selected layer
2. Which resource requirements can fewer trainable LoRA parameters reduce compared with full fine-tuning?
   • Every base-model computation by definition
   • All activation storage regardless of the training procedure
   • A guaranteed universal percentage of total runtime memory
   • Optimizer-state and task-checkpoint requirements
3. Co kontroluje ranga „r” w adapterze LoRA?
   • Wewnętrzny wymiar dwóch macierzy niskiego rzędu, określający pojemność adaptera
   • Liczba warstw w modelu podstawowym
   • Szybkość uczenia się podczas szkolenia
   • Ile procesorów graficznych jest wymaganych
4. Which storage and adaptation choices are combined in the QLoRA paper?
   • Updating every base weight directly as a 4-bit integer gradient
   • A frozen low-bit base with trainable LoRA adapters
   • Removing all adapters and all task data
   • A guarantee that every model fits every consumer GPU
5. What can a compatible dense merge of a LoRA update remove from that layer's inference computation?
   • The need to evaluate the base layer at all
   • Every possible dtype compatibility requirement
   • A requirement that all future task updates become identical
   • The need to evaluate a separate adapter branch
6. Which reuse pattern is supported by storing compatible task adapters separately from a base checkpoint?
   • Every adapter automatically makes every task factually correct
   • All serving systems must switch tasks with zero overhead
   • Multiple task updates can share the same pretrained base
   • The base checkpoint is no longer needed for inference
7. For a stated 4-by-4 weight update with rank-one factors shaped 4-by-1 and 1-by-4, how many factor parameters are there?
   • Szesnaście
   • Osiem
   • Cztery
   • Thirty-two
8. What does QLoRA's double-quantization technique compress in addition to the base weight representation?
   • The quantization constants
   • Every task label into a second unrelated category
   • The adapter gradients twice by definition
   • A second complete copy of the base model