Did you know that every time you download a language model to your computer, you're downloading powerful compression technology as well?

gptzip is a python library that uses pre-trained language models as string compressors. It's compatible out-of-the-box with language models from HuggingFace transformers and uses arithmetic coding (which is theoretically optimal) to compress strings based on language model probability distributions.

This all works because of Shannon's source coding theorem which connects probability distributions and compression. Since language models like GPT-3 give us probabilities over strings, we can literally use them as compressors. gptzip makes this trivial.

pip install gptzip

You can use gptzip to check the number of bytes a language model requires to encode a string (to compare against e.g. gzip or the original byte count):

model = "gpt2"
lm = transformers.AutoModelForCausalLM.from_pretrained(model)
tokenizer = transformers.AutoTokenizer.from_pretrained(model)
string = "Sailing on the seven seas"
coder = gptzip.ArithmeticCoder(lm=lm, tokenizer=tokenizer)
code, num_padded_bits = coder.encode(
    string, 
    return_num_padded_bits=True, 
)
assert len(code) == 5

Perhaps even more useful is to use gptzip as a true file compressor. In this case, code

model = "gpt2"
lm = transformers.AutoModelForCausalLM.from_pretrained(model)
tokenizer = transformers.AutoTokenizer.from_pretrained(model)
string = "How much would could a woodchuck chuck?"
coder = gptzip.ArithmeticCoder(lm=lm, tokenizer=tokenizer)
code, num_padded_bits = coder.encode(
    string, 
    return_num_padded_bits=True, 
)
print(f"Code: {to_binary(code)} ({len(code)} bytes)")
decoded_string = coder.decode(code, num_padded_bits=num_padded_bits)
assert decoded_string == string

Some features that would be nice to add:

• Other compression techniques such as Huffman
• Benchmarking against other compressions and add numbers to README
• Support for other language modeling softwares such as VLLM
• Compress multiple strings in batch

Thanks to DeepMind implementation for helping me implement Arithmetic coding in Python. I learned a lot from their implementation and paper, Language Modeling Is Compression.

I also am indebted to Mark Nelson for his incredibly blog post Data Compression With Arithmetic Coding. It was invaluable for me while learning about this topic, especially the lossless implementation of arithmetic coding using binary fractions. It's one of the best blog posts that I have ever read.