Idanwo ti o sopọ mọ itọsọna · Alabọde Ipele

Awoṣe Awoṣe Gigun

Review context windows, attention costs, position-sensitive retrieval, FlashAttention and evaluations of effective context use.

Jẹmọ itọsọna awọn ọnaAwoṣe Ọrọ Ọrọ Gigun
Ibeere 1 ti 8

Kini 'Frese ọrọ ọrọ' awoṣe kan tọka si?

Awotẹlẹ gbogbo ibeere8 ibeere

Awọn idahun ati awọn alaye han bi o ṣe n mu ibeere naa.

  1. Kini 'Frese ọrọ ọrọ' awoṣe kan tọka si?
    • Nọmba ti o pọ julọ ti awọn ami-ami ti o le lọ si ni iwe-iwọle kan
    • Iwọn iboju ti wiwo iwiregbe
    • Nọmba awọn olumulo laaye ni ẹẹkan
    • Bi o gun awoṣe ti a ti oṣiṣẹ
  2. Kini idi ti akiyesi ara ẹni boṣewa di gbowolori fun awọn igbewọle gigun?
    • O gbooro laini ati pe o jẹ olowo poku nigbagbogbo
    • Iṣiro rẹ ati iranti dagba ni ẹẹmẹrin pẹlu gigun ọkọọkan
    • Ko le ka ju aami kan lọ
    • O ṣiṣẹ nikan lori awọn aworan
  3. Kini iṣoro 'ti sọnu ni aarin'?
    • Awọn awoṣe gbagbe ikẹkọ wọn patapata
    • Awọn awoṣe le ka arin iwe nikan
    • Awọn awoṣe ṣe iranti alaye ni ibẹrẹ ati opin titẹ sii gigun ti o dara ju awọn ododo ti a sin ni aarin
    • Awọn awoṣe tumọ aarin si ede miiran
  4. Kini FlashAttention ni akọkọ ṣe ilọsiwaju?
    • O mu ki awọn awoṣe o wu ọrọ to gun
    • O ṣe afikun data ikẹkọ diẹ sii laifọwọyi
    • O yọ awọn nilo fun eyikeyi akiyesi
    • O dinku igo iranti pẹlu IO-mọ, iṣiro tile
  5. Kini ipa wo ni awọn ifibọ ipo rotari (RoPE) ṣe ninu awọn awoṣe ọrọ-gun?
    • Wọn ṣe koodu awọn ipo aami ati, pẹlu interpolation, ṣe iranlọwọ fun gbogbogbo si awọn ilana to gun
    • Wọn pa awọn ami-ami kuro lati inu titẹ sii
    • Wọn tumọ ọrọ laarin awọn ede
    • Wọn tọju awọn iwuwo awoṣe
  6. What does prompt-cache research reuse across compatible requests?
    • The final answer regardless of the new query
    • Computed attention state for reusable prompt segments
    • A fixed set of token IDs without model computation
    • The model’s training data
  7. What does the RULER benchmark add beyond locating one isolated “needle” fact?
    • Iyasọtọ aworan nikan
    • Audio transcription from waveforms
    • Tasks such as multi-hop tracing and aggregation over long contexts
    • GPU memory allocation benchmarks
  8. Why test several task types when measuring a model’s effective context length?
    • Every task uses the same computation
    • The advertised window guarantees all tasks work
    • Task type never affects context use
    • A model may handle simple retrieval yet fail to combine or track information across a long input