Language Model Zoo

⚠️⚠️⚠️ This project is no longer actively maintained by the Computational Psycholinguistics Laboratory. ⚠️⚠️⚠️

We do not guarantee the functionality or accuracy of the LM Zoo framework — use at your own risk!

You may be interested in the following active projects (as of June 2023):

• minicons enables easy Python access to neural network language model representations and probability/surprisal estimates.
• the Brain Score Language project provides tools for extracting behavioral and representational quantities from computational language models, and many benchmarks for evaluating the human-likeness of these models
• an experimental SyntaxGym implementation built directly into the Huggingface evaluate framework

---

The Language Model Zoo is an open-source repository of state-of-the-art language models, designed to support black-box access to model predictions and representations. It provides the command line tool lm-zoo, a standard interface for interacting with language models.

You can use lm-zoo to

1. compute language model predictions at the word level,
2. extract token-level surprisal data (popularly used in psycholinguistic experiments), and
3. preprocess corpora according to a language model's particular tokenization standards.

Quick links:

• Quickstart
• Supported models
• Contributing models

Getting started

Running language models from this repository requires Docker.

You can install the lm-zoo via pip:

$ pip install lm-zoo

List available language models:

$ lm-zoo list
gpt2
        Image URI:  docker.io/cpllab/language-models:gpt2
        Full name: None
        Reference URL: https://openai.com/blog/better-language-models/
        Maintainer: None
        Last updated: None
RNNG
        Image URI:  docker.io/cpllab/language-models:rnng
        Full name: None
        Reference URL: TODO
        Maintainer: None
        Last updated: None
ordered-neurons
        Image URI:  docker.io/cpllab/language-models:ordered-neurons
        Full name: None
        Reference URL: https://github.com/yikangshen/Ordered-Neurons
        Maintainer: None
        Last updated: None
...

Tokenize some text according to a language model's standard:

$ wget https://cpllab.github.io/lm-zoo/metamorphosis.txt -O metamorphosis.txt
$ lm-zoo tokenize gpt2 metamorphosis.txt
Pulling latest Docker image for cpllab/language-models:gpt2.
One Ġmorning , Ġwhen ĠGreg or ĠSam sa Ġwoke Ġfrom Ġtroubled Ġdreams , Ġhe Ġfound Ġhimself Ġtransformed Ġin Ġhis Ġbed Ġinto Ġa Ġhorrible Ġver min .
He Ġlay Ġon Ġhis Ġarmour - like Ġback , Ġand Ġif Ġhe Ġlifted Ġhis Ġhead Ġa Ġlittle Ġhe Ġcould Ġsee Ġhis Ġbrown Ġbelly , Ġslightly Ġdom ed Ġand Ġdivided Ġby Ġar ches Ġinto Ġstiff Ġsections .
The Ġbed ding Ġwas Ġhardly Ġable Ġto Ġcover Ġit Ġand Ġseemed Ġready Ġto Ġslide Ġoff Ġany Ġmoment .
...

Get token-level surprisals for text data:

$ lm-zoo get-surprisals ngram metamorphosis.txt
sentence_id     token_id        token   surprisal
1       1       one     7.76847
1       2       morning 9.40638
1       3       ,       1.05009
1       4       when    7.08489
1       5       gregor  18.8963
1       6       <unk>   4.27466
1       7       woke    19.0607
1       8       from    10.3404
1       9       troubled        17.478
1       10      dreams  10.671
1       11      ,       3.39374
1       12      he      5.99193
1       13      found   8.07358
1       14      himself 2.92718
1       15      transformed     16.7328
1       16      in      5.32057
1       17      his     7.26454
1       18      bed     9.78166
1       19      into    8.90954
1       20      a       3.72355
1       21      horrible        14.2477
1       22      <unk>   3.56907
1       23      .       3.90242
1       24      </s>    22.8395
2       1       he      4.43708
2       2       lay     14.1721
...

For more information, see our Quickstart tutorial.