perplexed

This library is based on the idea from Andrej Karpathy on understanding the failure cases of a model by looking at the worst predictions. Specifically, this library focuses on calculating the perplexity of Large Language Models (LLMs) such as GPT-2 and BERT. The idea is to calculate the perplexity of a model on a dataset at the per token level. This allows us to understand where the model is perplexed and where it is not. This is useful for debugging and understanding the model.

Install

pip install perplexed

How to use

Using the API

perplexed is designed to work with the HuggingFace ecosystem and is built on top of the transformers and datasets libraries. The API is designed to be simple and easy to use. The main function is perplexed which takes in a model, dataset, and tokenizer and returns a simple Counter object with the perplexity of each token in the dataset. Here is an example of how to use it:

from perplexed.core import perplexed

tokenizer = AutoTokenizer.from_pretrained("EleutherAI/gpt-neo-125M")
tokenizer.pad_token = tokenizer.eos_token
model = AutoModelForCausalLM.from_pretrained("EleutherAI/gpt-neo-125M")

dataset = load_dataset("wikitext", "wikitext-2-raw-v1", split="test").select(range(100))
# filter out empty strings
dataset = dataset.filter(lambda x: len(x["text"]) > 0)

perplexity_cnt = perplexed(model, dataset, tokenizer=tokenizer, column="text", batch_size=1, device="cpu")
perplexity_cnt.most_common(10)

Found cached dataset wikitext (/home/nathan/.cache/huggingface/datasets/wikitext/wikitext-2-raw-v1/1.0.0/a241db52902eaf2c6aa732210bead40c090019a499ceb13bcbfa3f8ab646a126)
Loading cached processed dataset at /home/nathan/.cache/huggingface/datasets/wikitext/wikitext-2-raw-v1/1.0.0/a241db52902eaf2c6aa732210bead40c090019a499ceb13bcbfa3f8ab646a126/cache-68eb731029328d8b.arrow
Loading cached processed dataset at /home/nathan/.cache/huggingface/datasets/wikitext/wikitext-2-raw-v1/1.0.0/a241db52902eaf2c6aa732210bead40c090019a499ceb13bcbfa3f8ab646a126/cache-1c1cd85efcee4db8.arrow

[(' wired', 60983688.0),
 (' 768', 21569838.0),
 (' shatter', 12281687.0),
 (' unsett', 8289435.0),
 (' ignited', 6605209.0),
 (' Tanz', 4834899.0),
 (' Influence', 4153321.75),
 (' Career', 4064189.0),
 (' Television', 2325870.75),
 (' Moral', 2243574.5)]