Impara Evaluating Language Models with Perplexity

Scorri per mostrare il menu

Perplexity is the standard metric for evaluating language model quality during pre-training. It measures how uncertain the model is when predicting the next token – a lower value means the model assigns higher probability to the correct tokens.

The Formula

For a sequence of $N$ tokens with predicted probabilities $\ p_1, p_2, \ldots, p_N\$ for the correct token at each step:

PP = \exp\left(\frac{1}{N} \sum_{i=1}^{N} -\log p_i\right)

This is simply the exponentiation of the average cross-entropy loss. If your training loop already computes cross-entropy loss, perplexity is just torch.exp(loss).

Computing Perplexity in PyTorch


              123456789101112131415161718
            
import torch
import torch.nn.functional as F

# Simulating model logits and targets for a single batch
vocab_size = 1000
seq_len = 20
batch_size = 4

logits = torch.rand(batch_size, seq_len, vocab_size)
targets = torch.randint(0, vocab_size, (batch_size, seq_len))

# Cross-entropy loss (mean over all tokens)
loss = F.cross_entropy(logits.view(-1, vocab_size), targets.view(-1))

# Perplexity
perplexity = torch.exp(loss)
print(f"Loss: {loss.item():.4f}")
print(f"Perplexity: {perplexity.item():.2f}")

A randomly initialized model over a vocabulary of 1000 tokens should produce a perplexity close to 1000 – it is essentially guessing uniformly. As training progresses, perplexity drops.

Run this locally and try replacing the random logits with a uniform distribution to verify the relationship between vocabulary size and initial perplexity.

Strengths and Limitations

Perplexity is easy to compute, interpretable, and consistent across training runs on the same dataset. However, it has real limitations:

it is sensitive to tokenization – comparing perplexity across models with different vocabularies is not meaningful;
it does not capture output quality for generation tasks – a model can have low perplexity while producing repetitive or incoherent text;
it only reflects next-token prediction accuracy, not downstream task performance.

Tutto è chiaro?

Grazie per i tuoi commenti!

Sezione 1. Capitolo 10

Chieda ad AI

Chieda pure quello che desidera o provi una delle domande suggerite per iniziare la nostra conversazione

Sezione 1. Capitolo 10