Multiple Choice

Pick the H100 workload where spot pricing is genuinely safe to deploy on

Which quantization combo squeezes the most decode bandwidth per percent of quality lost?

A response comes back with finish_reason = 'length' which interpretation is correct?

Pick Claude Sonnet 4.6 over Opus 4.7, when is that the right call?

Pick the workload where a provider's Batch API beats the realtime endpoint

Why does completion-only loss usually outperform full-sequence loss on the same SFT data?

Pick the most likely cause when training loss drops for 1k steps, plateaus for 2k, then drops again at step 3k

LoRA freezes the base weights: does that mean activation checkpointing on the base is unnecessary?

Why does using the same model family as synthetic-data teacher AND eval judge inflate FT scores?

RLOO's name hides its baseline trick: what does 'Leave-One-Out' literally compute?

Same LoRA recipe on H100 vs A100: what precision should each team default to?

Only 200 examples for a niche format task: which PEFT method gives the best shot?

DDP vs FSDP for a 13B full fine-tune on 4xA100-40GB: which is feasible?

Pick the most accurate one-sentence summary of the tokenizer's role in fine-tuning.

Teacher forcing during SFT: which description nails it?

Should the system message contribute to the SFT loss?

ShareGPT format: what data shape does it use?

save_strategy='steps' vs save_strategy='epoch': what behaviour changes?

Roles in a chat training example: pick the three that are standard.

Reward model in RLHF: what does it score, and on what data does it train?

Pick the sensible `num_train_epochs` for a typical instruction-SFT run on a curated dataset

When does a PEFT config actually need `modules_to_save` set, and what goes in it?

Modern LLM fine-tuning defaults to causal LM or masked LM, pick one and say why

After calling `merge_and_unload()` on a PEFT model, what does the returned object look like?

Which loss function does supervised fine-tuning actually minimize?