Ikibazo gihuza ibibazo · Hagati Urwego

Amategeko yo gupima Chinchilla

Gerageza Chinchilla amahugurwa-yingengo yimari, 70B / 1.4T urugero, urugero rwo gupima, kubara hafi, no gutandukanya uburyo bwo kohereza.

Inzira zijyanye no kuyoboraAmategeko yo gupima Chinchilla
Ikibazo 1 Bya 8

What does Chinchilla's finding that parameters and tokens should scale in similar proportions mean?

Reba buri kibazo8 ibibazo

Ibisubizo nibisobanuro bigaragara mugihe ufata ikibazo.

  1. What does Chinchilla's finding that parameters and tokens should scale in similar proportions mean?
    • Both should grow with comparable scaling exponents under the studied training-compute objective
    • The numerical token count must equal the parameter count
    • Training data does not affect the fitted loss
    • Every deployment must choose the largest possible model
  2. What token-per-parameter ratio does the 70B-parameter, 1.4T-token Chinchilla experiment illustrate?
    • Exactly one token per parameter
    • About two thousand tokens per parameter
    • About twenty training tokens per parameter
    • A ratio determined only by the number of GPUs
  3. Ni ubuhe bwoko Chinchilla yarushije nubwo ari buto cyane?
    • GPT-2
    • Gopher
    • BERT
    • Mamba
  4. Chinchilla yashakaga kuvuga iki kubijyanye na moderi nka GPT-3 icyo gihe?
    • Bararengereye
    • Bafite compte nkeya cyane
    • Bakoresheje tokenizer itari yo
    • Bahawe imyitozo ugereranije nubunini bwabo
  5. Hafi ya compte yagereranijwe gute mu isesengura rya Chinchilla?
    • Ugereranije numubare wa GPU gusa
    • Yigenga yubunini bwamakuru
    • Ugereranije nibihe ibihe byamahugurwa ibimenyetso
    • Ugereranije nubunini bwamagambo
  6. Why might a deployment objective favor a smaller model trained on extra data rather than the training-compute optimum alone?
    • Repeated inference costs can matter alongside the initial training budget
    • Inference has no resource cost for any model
    • The Chinchilla study proved every scaling relationship false
    • More data always reduces both training and serving work to zero
  7. Under the approximation that training compute is proportional to N times D, how must D change if N doubles while that compute budget stays fixed?
    • D must also double
    • D must quadruple
    • D must remain unchanged
    • D must approximately halve
  8. Which three types of term appear in Chinchilla’s parametric loss fit?
    • Only GPU count, electricity price, and elapsed seconds
    • An irreducible base term, a model-size contribution, and a data-size contribution
    • Only vocabulary size, file size, and model name
    • Only a linear token count with no model-size term