Flashcard

Rank r in LoRA: what does the number actually count?

Preference pair: name the two pieces and what the label expresses.

PEFT: unpack the acronym and the core problem it tackles.

Paged optimizer in QLoRA, what gets paged, and to where?

`max_seq_len` controls one bound, what happens to examples that exceed it?

Why ramp the learning rate during warmup instead of starting at the peak value?

Gradient norm clipping, name the failure mode it exists to prevent

Gradient checkpointing, what does it trade, and roughly by how much?

Why does dropping the EOS token from SFT labels produce a model that never stops generating?

Name the signal early stopping watches, and explain the 'patience' knob

Explain what gets quantized the second time in QLoRA's double-quantization step

Define the role of a chat template when fine-tuning on (system, user, assistant) data

Micro batch vs effective batch in fine-tuning: give the relationship

lora_alpha in plain English: what role does that scalar play?

Alpaca-format examples carry three fields: name them

Accelerate launch: what does that one CLI actually wire up?

How does softmax turn an attention score into an attention weight?

Name the scalar that QK^T is divided by inside scaled dot-product attention.

RoPE, spell out the acronym and name exactly which vectors it rotates

Inside a transformer, what is the residual stream and how does attention interact with it?

Name what wraps the attention sub-layer in every transformer block.

Prefill and decode: name the two phases of LLM inference and say which one is compute-bound

What does a padding mask zero out, and why is it needed when batching variable-length sequences?

Describe the W_O projection in multi-head attention, its shape and what it mixes.

Which config token names the count of parallel attention heads in a layer?