Text classification with Python 3.14's ZSTD module
maxhalford.github.io190 points by alexmolas 3 days ago
190 points by alexmolas 3 days ago
This looks like a nice rundown of how to do this with Python's zstd module.
But, I'm skeptical of using compressors directly for ML/AI/etc. (yes, compression and intelligence are very closely related, but practical compressors and practical classifiers have different goals and different practical constraints).
Back in 2023, I wrote two blog-posts [0,1] that refused the results in the 2023 paper referenced here (bad implementation and bad data).
Concur. Zstandard is a good compressor, but it's not magical; comparing the compressed size of Zstd(A+B) to the common size of Zstd(A) + Zstd(B) is effectively just a complicated way of measuring how many words and phrases the two documents have in common. Which isn't entirely ineffective at judging whether they're about the same topic, but it's an unnecessarily complex and easily confused way of doing so.
I do not know inner details of Zstandard, but I would expect that it to least do suffix/prefix stats or word fragment stats, not just words and phrases.
Good on you for attempting to reproduce the results & writing it up, and reporting the issue to the authors.
> It turns out that the classification method used in their code looked at the test label as part of the decision method and thus led to an unfair comparison to the baseline results
This has been possible with the zlib module since 1997 [EDIT: zlib is from '97. The zdict param wasn't added until 2012]. You even get similar byte count outputs to the example and on my machine, it's about 10x faster to use zlib.
import zlib
input_text = b"I ordered three tacos with extra guacamole"
tacos = b"taco burrito tortilla salsa guacamole cilantro lime " * 50
taco_comp = zlib.compressobj(zdict=tacos)
print(len(taco_comp.compress(input_text) + taco_comp.flush()))
# prints 41
padel = b"racket court serve volley smash lob match game set " * 50
padel_comp = zlib.compressobj(zdict=padel)
print(len(padel_comp.compress(input_text) + padel_comp.flush()))
# prints 54True. The post calls out that “you have to recompress the training data for each test document” with zlib (otherwise input_text would taint it), but you can actually call Compress.copy().
zdict was added in Python 3.3, though, so it’s 2012, not 1997. (It might have worked before, just not a part of the official API :-)
Ah, okay. Didn't realize that. I used either zlib or gzip long, long ago but never messed with the `zdict` param. Thanks for pointing that out.
The application of compressors for text statistics is fun, but it's a software equivalent of discovering that speakers and microphones are in principle the same device.
(KL divergence of letter frequencies is the same thing as ratio of lengths of their Huffman-compressed bitstreams, but you don't need to do all this bit-twiddling for real just to count the letters)
The article views compression entirely through Python's limitations.
> gzip and LZW don’t support incremental compression
This may be true in the Python's APIs, but is not true about these algorithms in general.
They absolutely support incremental compression even in APIs of popular lower-level libraries.
Snapshotting/rewinding of the state isn't exposed usually (custom gzip dictionary is close enough in practice, but a dedicated API would reuse its internal caches). Algorithmically it is possible, and quite frequently used by the compressors themselves: Zopfli tries lots of what-if scenarios in a loop. Good LZW compression requires rewinding to a smaller symbol size and restarting compression from there after you notice the dictionary stopped being helpful. The bitstream has a dedicated code for this, so this isn't just possible, but baked into the design.
> The application of compressors for text statistics is fun, but it's a software equivalent of discovering that speakers and microphones are in principle the same device.
I think it makes sense to explore it from practical standpoint, too. It’s in Python stdlib, and works reasonably well, so for some applications it might be good enough.
It’s also fairly easy to implement in other languages with zstd bindings, or even shell scripts:
$ echo 'taco burrito tortilla salsa guacamole cilantro lime' > /tmp/tacos.txt
$ zstd --train $(yes '/tmp/tacos.txt' | head -n 50) -o tacos.dict
[...snip]
$ echo 'racket court serve volley smash lob match game set' > /tmp/padel.txt
$ zstd --train $(yes '/tmp/padel.txt' | head -n 50) -o padel.dict
[...snip]
$ echo 'I ordered three tacos with extra guacamole' | zstd -D tacos.dict | wc -c
57
$ echo 'I ordered three tacos with extra guacamole' | zstd -D padel.dict | wc -c
60Or with the newsgroup20 dataset:
curl http://qwone.com/~jason/20Newsgroups/20news-19997.tar.gz | tar -xzf -
cd 20_newsgroups
for f in *; do zstd --train "$f/*" -o "../$f.dict"; done
cd ..
for d in *.dict; do
cat 20_newsgroups/misc.forsale/74150 | zstd -D "$d" | wc -c | tr -d '\n'; echo " $d";
done | sort | head -n 3
422 misc.forsale.dict
462 rec.autos.dict
463 comp.sys.mac.hardware.dict
Great overview. In 2023 I wrote about classifying political emails with Zstd.¹
¹ https://matthodges.com/posts/2023-10-01-BIDEN-binary-inferen...
Ooh, totally. Many years ago I was doing some analysis of parking ticket data using gnuplot and had it output a chart png per-street. Not great, but worked well to get to the next step of that project of sorting the directory by file size. The most dynamic streets were the largest files by far.
Another way I've used image compression to identify cops that cover their body cameras while recording -- the filesize to length ratio reflects not much activity going on.
The stdlib inclusion angle is what makes this interesting to me beyond the compression-as-classifier debate.
Before 3.14, doing this required pip installing pyzstd or python-zstandard, which meant an extra dependency and a C extension build step. Having zstd in the stdlib means you can do dictionary-based compression tricks in any Python 3.14 environment without worrying about whether the deployment target has build tools or whether the package version matches across your team.
That matters more than it sounds for reproducibility. The gzip/zlib approach kenschutte mentions has been in stdlib forever, but zstd dictionaries are meaningfully better at learning domain-specific patterns from small training corpora. The difference between 41 and 54 bytes in the taco example is small, but on real classification tasks with thousands of categories the gap compounds.
Python 3.14 quietly shipped a lot of practical stuff like this that gets overshadowed by the free-threading and JIT headlines. The ZSTD module, t-strings for template injection safety, and deferred annotation evaluation are all things that change how you write everyday code, not just how the runtime performs.
Is this an AI response? This account was created 4 days ago and all its comments follow the exact same structure. The comments are surprisingly not easy to tell it's AI but it always makes sure to include a "it's X, not Y" conclusion.
There's also Normalized Google Distance (a distance metric using the number of search results as a proxy), which can be used for text classification.
Python's zlib does support incremental compression with the zdict parameter. gzip has something similar but you have to do some hacky thing to get at it since the regular Python API doesn't expose the entry point. I did manage to use it from Python a while back, but my memory is hazy about how I got to it. The entry point may have been exposed in the code module but undocumented in the Python manual.
The speed comparison is weird.
The author sets the solver to saga, doesn’t standardize the features, and uses a very high max_iter.
Logistic Regression takes longer to converge when features are not standardized.
Also, the zstd classifier time complexity scales linearly with the number of classes, logistic regression doesn’t. You have 20 (it’s in the name of the dataset), so why only use 4.
It’s a cool exploration of zstd. But please give the baseline some love. Not everything has to be better than something to be interesting.
In my PhD more than a decade ago, I ended up using png image file sizes to classify different output states from simulations of a system under different conditions. Because of the compressions, homogenous states led to much smaller file size than the heterogenous states. It was super super reliable.
Why did python include ZSTD? are people passing around files compressed with this algorithm? It's the first I've ever heard of it.
So that’s why Facebook developed a ”compression” library that has snaked itself into various environments.
Sweet! I love clever information theory things like that.
It goes the other way too. Given that LLMs are just lossless compression machines, I do sometimes wonder how much better they are at compressing plain text compared to zstd or similar. Should be easy to calculate...
EDIT: lossless when they're used as the probability estimator and paired with something like an arithmetic coder.
> Given that LLMs are just lossless compression machines, I do sometimes wonder how much better they are at compressing plain text compared to zstd or similar. Should be easy to calculate...
The current leader on the Hutter Prize (http://prize.hutter1.net/) are all LLM based.
It can (slowly!!) compress a 1GB dump of Wikipedia to 106Mb
By comparison GZip can compress it to 321Mb
See https://mattmahoney.net/dc/text.html for the current leaderboard
I've actually been experimenting with that lately. I did a really naive version that tokenizes the input, feeds the max context window up to the token being encoded into an LLM, and uses that to produce a distribution of likely next tokens, then encodes the actual token with Huffman Coding with the LLM's estimated distribution. I could get better results with arithmetic encoding almost certainly.
It outperforms zstd by a long shot (I haven't dedicated the compute horsepower to figuring out what "a long shot" means quantitatively with reasonably small confidence intervals) on natural language, like wikipedia articles or markdown documents, but (using GPT-2) it's about as good as zstd or worse than zstd on things like files in the Kubernetes source repository.
You already get a significant amount of compression just out of the tokenization in some cases ("The quick red fox jumps over the lazy brown dog." encodes to one token per word plus one token for the '.' for the GPT-2 tokenizer), where as with code a lot of your tokens will just represent a single character so the entropy coding is doing all the work, which means your compression is only as good as the accuracy of your LLM, plus the efficiency of your entropy coding.
I would need to be encoding multiple tokens per "word" with Huffman Coding to hit the entropy bounds, since it has a minimum of one bit per character, so if tokens are mostly just one byte then I can't do better than a 12.5% compression ratio with one token per word. And doing otherwise gets computationally infeasible very fast. Arithmetic coding would do much better especially on code because it can encode a word with fractional bits.
I used Huffman coding for my first attempt because it's easier to implement and most libraries don't support dynamically updating the distribution throughout the process.
I do not agree on the "lossless" adjective. And even if it is lossless, for sure it is not deterministic.
For example I would not want a zip of an encyclopedia that uncompresses to unverified, approximate and sometimes even wrong text. According to this site : https://www.wikiwand.com/en/articles/Size%20of%20Wikipedia a compressed Wikipedia without medias, just text is ~24GB. What's the medium size of an LLM, 10 GB ? 50 GB ? 100 GB ? Even if it's less, it's not an accurate and deterministic way to compress text.
Yeah, pretty easy to calculate...
(to be clear this is not me arguing for any particular merits of llm-based compression, but) you appear to have conflated one particular nondeterministic llm-based compression scheme that you imagined with all possible such schemes, many of which would easily fit any reasonable definitions of lossless and deterministic by losslessly doing deterministic things using the probability distributions output by an llm at each step along the input sequence to be compressed.
With a temperature of zero, LLM output will always be the same. Then it becomes a matter of getting it to output the exact replica of the input: if we can do that, it will always produce it, and the fact it can also be used as a bullshit machine becomes irrelevant.
With the usual interface it’s probably inefficient: giving just a prompt alone might not produce the output we need, or it might be larger than the thing we’re trying to compress. However, if we also steer the decisions along the way, we can probably give a small prompt that gets the LLM going, and tweak its decision process to get the tokens we want. We can then store those changes alongside the prompt. (This is a very hand-wavy concept, I know.)
There's an easier and more effective way of doing that - instead of trying to give the model an extrinsic prompt which makes it respond with your text, you use the text as input and, for each token, encode the rank of the actual token within the set of tokens that the model could have produced at that point. (Or an escape code for tokens which were completely unexpected.) If you're feeling really crafty, you can even use arithmetic coding based on the probabilities of each token, so that encoding high-probability tokens uses fewer bits.
From what I understand, this is essentially how ts_zip (linked elsewhere) works.
> With a temperature of zero, LLM output will always be the same
Ignoring GPU indeterminism, if you are running a local LLM and control batching, yes.
If you are computing via API / on the cloud, and so being batched with other computations, then no (https://thinkingmachines.ai/blog/defeating-nondeterminism-in...).
But, yes, there is a lot of potential from semantic compression via AI models here, if we just make the efforts.
Compression can be generalized as probability modelling (prediction) + entropy coding of the difference between the prediction and the data. The entropy coding has known optimal solutions.
So yes, LLMs are nearly ideal text compressors, except for all the practical inconveniences of their size and speed (they can be reliably deterministic if you sacrifice parallel execution and some optimizations).
Aren't LLMs lossy? You could make them lossless by also encoding a diff of the predicted output vs the actual text.
Edit to soften a claim I didn't mean to make.
LLMs are good at predicting the next token. Basically you use them to predict what are the probabilities of the next tokens to be a, b, or c. And then use arithmetic coding to store which one matched. So the LLM is used during compression and decompression.
Yes LLMs are always lossy, unless their size / capacity is so huge they can memorize all their inputs. Even if LLMs were not resource-constrained, one would expect lossy compression due to batching and the math of the loss function. Training is such that it is always better for the model to accurately approximate the majority of texts than to approximate any single text with maximum accuracy.
So you just discovered pca in some other form?
Can `copy.deepcopy` not help?
[dead]