Language Model Perplexity (LM-PPL)

Perplexity measures how predictable a text is by a language model (LM), and it is often used to evaluate fluency or proto-typicality of the text (lower the perplexity is, more fluent or proto-typical the text is). LM-PPL is a python library to calculate perplexity on a text with any types of pre-trained LMs. We compute an ordinary perplexity for recurrent LMs such as GPT3 (Brown et al., 2020) and the perplexity of the decoder for encoder-decoder LMs such as BART (Lewis et al., 2020) or T5 (Raffel et al., 2020), while we compute pseudo-perplexity (Wang and Cho, 2018) for masked LMs.

Get Started

Install via pip.

pip install lmppl

Example

Let's solve sentiment analysis with perplexity as an example! Remember the text with lower perplexity is better, so we compare two texts (positive and negative) and choose the one with lower perplexity as the model prediction.

1. Recurrent LM including variants of GPT.

import lmppl

scorer = lmppl.LM('gpt2')
text = [
    'sentiment classification: I dropped my laptop on my knee, and someone stole my coffee. I am happy.',
    'sentiment classification: I dropped my laptop on my knee, and someone stole my coffee. I am sad.'
]
ppl = scorer.get_perplexity(text)
print(list(zip(text, ppl)))
>>> [
  ('sentiment classification: I dropped my laptop on my knee, and someone stole my coffee. I am happy.', 136.64255272925908),
  ('sentiment classification: I dropped my laptop on my knee, and someone stole my coffee. I am sad.', 139.2400838400971)
]
print(f"prediction: {text[ppl.index(min(ppl))]}")
>>> "prediction: sentiment classification: I dropped my laptop on my knee, and someone stole my coffee. I am happy."

2. Masked LM including variants of BERT.

import lmppl

scorer = lmppl.MaskedLM('microsoft/deberta-v3-small')
text = [
    'sentiment classification: I dropped my laptop on my knee, and someone stole my coffee. I am happy.',
    'sentiment classification: I dropped my laptop on my knee, and someone stole my coffee. I am sad.'
]
ppl = scorer.get_perplexity(text)
print(list(zip(text, ppl)))
>>> [
  ('sentiment classification: I dropped my laptop on my knee, and someone stole my coffee. I am happy.', 1190212.1699246117),
  ('sentiment classification: I dropped my laptop on my knee, and someone stole my coffee. I am sad.', 1152767.482071837)
]
print(f"prediction: {text[ppl.index(min(ppl))]}")
>>> "prediction: sentiment classification: I dropped my laptop on my knee, and someone stole my coffee. I am sad."

3. Encoder-Decoder LM including variants of T5 and BART.

import lmppl

scorer = lmppl.EncoderDecoderLM('google/flan-t5-small')
inputs = [
    'sentiment classification: I dropped my laptop on my knee, and someone stole my coffee.',
    'sentiment classification: I dropped my laptop on my knee, and someone stole my coffee.'
]
outputs = [
    'I am happy.',
    'I am sad.'
]
ppl = scorer.get_perplexity(input_texts=inputs, output_texts=outputs)
print(list(zip(outputs, ppl)))
>>> [
  ('I am happy.', 4138.748977714201),
  ('I am sad.', 2991.629250051472)
]
print(f"prediction: {outputs[ppl.index(min(ppl))]}")
>>> "prediction: I am sad."

Models

Below are some examples of popular models and the corresponding model type to use within the lmppl package.

Model	HuggingFace ID	Model Type
BERT	google-bert/bert-base-uncased	MaskedLM
Roberta	roberta-large	MaskedLM
GPT 2	gpt2-xl	LM
flan-ul2	google/flan-ul2	EncoderDecoderLM
GPT-NeoX	EleutherAI/gpt-neox-20b	LM
OPT	facebook/opt-30b	LM
Mixtral	mistralai/Mixtral-8x22B-v0.1	LM
Llama 3	meta-llama/Meta-Llama-3-8B	LM

Tips

• Max Token Length: Each LM has its own max-token length (max_length for recurrent/masked LMs, and max_length_encoder and max_length_decoder for encoder-decoder LMs). Limiting those max-token will reduce the time to process the text, but it may affect the accuracy of the perplexity, so please experiment on your texts and decide an optimal token length.

• Batch Size: One can pass batch size to the function get_perplexity (eg. get_perplexity(text, batch_size=32)). As default, it will process all the text once, that may cause memory error if the number of texts is too large.