Ikibazo gihuza ibibazo · Hagati Urwego
Perplexity n'ikibazo cyo gupima ururimi
Ongera usubiremo ibimenyetso-byerekana ko bishoboka, kwambukiranya-entropiya, ibice bimwe, kugereranya kugereranya, BLEU, imipaka yubuziranenge, imiterere ya Windows, hamwe na masike yerekana.
Ikibazo 1 Bya 8
Under the same autoregressive text-scoring setup, what does lower perplexity indicate?
Reba buri kibazo8 ibibazo
Ibisubizo nibisobanuro bigaragara mugihe ufata ikibazo.
- Under the same autoregressive text-scoring setup, what does lower perplexity indicate?
- Lower average negative log-likelihood for the reference tokens
- A guarantee of more factually correct generated answers
- A guarantee of a larger parameter count
- A requirement that every top-choice prediction be correct
- Using natural-log negative log-likelihood, how is autoregressive perplexity computed from its average?
- By multiplying the parameter count by the context length
- By subtracting the average from the number of attention heads
- By taking the average learning rate
- By exponentiating the average negative log-likelihood
- Umunyamideli atanga amahirwe amwe mumagambo 10,000-yamagambo atize. Ni ubuhe buryo butangaje?
- 1
- 10,000
- 0
- 100
- Why do tokenizer and corpus settings matter when directly comparing per-token perplexity values?
- Perplexity ignores every scored reference token
- All corpora necessarily produce the same likelihood
- They change the units or reference data over which probabilities are averaged
- Only model parameter count determines the score
- Which translation-reference metric uses modified n-gram precision and a brevity penalty?
- Perplexity
- Kugereranya ikosa rya pigiseli ikosa
- Icyitegererezo cyo kubara
- BLEU
- Ni izihe mbogamizi zingenzi zisunika abashakashatsi kurenza urujijo rwo gusuzuma ubuziranenge bwisi?
- Perplexity ipima ijambo rikurikira-guhanura, ntabwo niba ibisohoka ari ingirakamaro, ukuri, cyangwa bifasha
- Perplexity bihenze cyane kubara
- Perplexity ntishobora kubarwa kumibare yinyandiko
- Perplexity ikora gusa mugihe cyo gufata umwanzuro, ntabwo ari imyitozo
- When scoring a long text with a fixed-context model, what can a sliding-window procedure provide that disjoint chunks lose near boundaries?
- A new vocabulary learned automatically at every boundary
- A guarantee that generated answers are factually correct
- More preceding context for the scored target tokens
- A requirement to count each overlapping target repeatedly
- Why is conventional left-to-right perplexity not directly the same score for a BERT-style masked language model?
- BERT has no learned parameters
- The masked model does not directly define the same autoregressive sequence factorization
- Every masked-token probability is necessarily zero
- Perplexity is defined only for image pixels