Idanwo ti o sopọ mọ itọsọna · Alabọde Ipele

EleutherAI adanwo

Review EleutherAI's open language models, Pile dataset, evaluation tools, and practical architecture choices such as RoPE.

Jẹmọ itọsọna awọn ọnaEleutheri
Ibeere 1 ti 8

How did EleutherAI first bring together its research community?

Awotẹlẹ gbogbo ibeere8 ibeere

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

  1. How did EleutherAI first bring together its research community?
    • As a Discord server for discussing GPT-3 and open language modeling
    • Gẹgẹbi ibẹrẹ iṣowo-owo ni Silicon Valley
    • Bi awọn kan ijoba iwadi yàrá
    • As a university’s official AI department
  2. Kini 'The Pile'?
    • Ọja chatbot ti ara ẹni
    • 825GB ṣisilẹ data ọrọ ti o ṣi silẹ fun awọn awoṣe ede ikẹkọ
    • A iru ti nkankikan Layer Layer
    • EleutherAI ká awọsanma iširo Syeed
  3. Ewo ninu iwọnyi jẹ awoṣe ede EleutherAI?
    • BERT-tobi
    • LLMA-2
    • GPT-NeoX-20B
    • PaLM
  4. Kini ijanu Igbelewọn LM ti a lo fun?
    • Ṣiṣẹda awọn aworan lati ọrọ
    • Compressing awoṣe òṣuwọn
    • Awọn awoṣe alejo gbigba bi API wẹẹbu
    • Awọn awoṣe ede Benchmarking kọja awọn iṣẹ ṣiṣe lọpọlọpọ
  5. Which organizations are named in EleutherAI’s retrospective as research supporters?
    • CoreWeave and Google TPU Research Cloud
    • Nvidia and the National Science Foundation only
    • The U.S. Department of Energy and NASA
    • Amazon and the European Space Agency
  6. Iyipada ajo wo ni EleutherAI ṣe ni ọdun 2023?
    • O ti gba nipasẹ OpenAI
    • O ṣe agbekalẹ gẹgẹbi ile-iṣẹ iwadii ti ko ni ere
    • O di owo hejii fun-èrè
    • O tiipa patapata
  7. Which positional method does the GPT-NeoX-20B paper describe using instead of learned absolute position embeddings?
    • A separate learned embedding for every document
    • A convolutional image-position map
    • Rotary Positional Embeddings (RoPE)
    • A random position permutation at each layer
  8. What architectural arrangement is highlighted as a GPT-NeoX-20B difference from GPT-3?
    • The model replaces its Transformer with a decision tree
    • Every layer removes the attention mechanism
    • It uses only an encoder and no autoregressive decoder
    • Attention and feed-forward layers can be computed in parallel