Large Language Models and the Argument from the Poverty of the Stimulus

Code and data for Large Language Models and the Argument from the Poverty of the Stimulus.

Getting started

• Install Python 3.9 and required packages:

pip install -r requirements.txt

Usage examples:

• Plotting surprisal curves for one or more models:

import viz

models = ("gpt2", "gpt3")
sentences = (
    "I know who you talked with yesterday",  # Grammatical.
    "I know who you talked with _ Mary _ yesterday",    # Ungrammatical.   
)

labels = ("Grammatical", "Ungrammatical")
viz.plot_surprisals(sentences, models=models, labels=labels)

Result:

Running filler/gap experiments

The grammars for the tasks described in the paper are defined in grammars.toml.

Running run_phenomenon.py runs all sentences generated from the grammar, for all combinations of [±filler, ±gap].

To get surprisal values for all four conditions on specific models, and calculate the interaction values:

python run_phenomenon.py --phenomenon {"ATB", "PG"} --models <comma separated list of models> --names {'tiny', 'small', 'large'}   

Supported models: bert-base-uncased, bert-large-uncased, gpt2, gpt3, grnn, childes_transformer_<i>, childes_lstm_<i>.

To run the experiments from the paper, for example for gpt2 and gpt3, run:

python run_phenomenon.py --phenomenon ATB --models gpt2,gpt3 --names tiny
python run_phenomenon.py --phenomenon PG --models gpt2,gpt3 --names small

Results are saved to data, for example:

• ATB__model__gpt3__cfg_tiny_v2023-10-30a.{csv,pickle} – surprisal values for each sentence generated by the grammar.
• ATB__model__gpt3__cfg_tiny_v2023-10-30a_full_interaction.{csv,pickle} – surprisal interaction values for each tuple of four [±filler, ±gap] sentences.

Paper data

• To reproduce the plots and stats reported in the paper, see notebooks in notebooks.

• Logs for the Transformer retraining task are in data/logs.zip.

• For the full retrained Transformer checkpoints, write to nur.lan@ens.psl.eu.

Raw surprisal values

All the sentences and raw surprisal values for the experiments reported in the paper are found under data/model_surprisals.zip.

To prevent test set contamination by large language models who train on crawled data and then test on it, all dataset files are zipped and password-protected. The password to all zip files is llm_pos.

See Jacovi et al., 2022 – Practical Strategies for Mitigating Data Contamination by Evaluation Benchmarks.

Model setup

• To use GPT-3, add your OpenAI API key as an environment variable:

export OPENAI_KEY="..."

lm-zoo models

The models GPT-2, GRNN, and JRNN are run using Docker based on the lm-zoo project.

Pull the following images to be able to get inferences from GPT2, GRNN, JRNN:

cpllab/language-models:gpt2
cpllab/language-models:jrnn
cpllab/language-models:grnn

• ! JRNN is extremely slow and takes 1-2 minutes per one sentence inference.

CHILDES models

Run ./get_childes_models.sh to download the CHILDES LSTM/Transformer models from Yedetore et al. (2023).

Running inference on CHILDES models:

Clone original repo:

git clone https://github.com/adityayedetore/lm-povstim-with-childes/

And add its location to PYTHONPATH:

export PYTHONPATH="$PYTHONPATH:/PATH/TO/lm-povstim-with-childes/src"

Retraining task

Getting the original English Wikipedia training data:

WIKI_DATA_DIR="./data/retraining/wikipedia"
mkdir -p $WIKI_DATA_DIR

wget -x https://dl.fbaipublicfiles.com/colorless-green-rnns/training-data/English/train.txt -O $WIKI_DATA_DIR/train.txt.original
wget -x https://dl.fbaipublicfiles.com/colorless-green-rnns/training-data/English/test.txt -O $WIKI_DATA_DIR/test.txt.original
wget -x https://dl.fbaipublicfiles.com/colorless-green-rnns/training-data/English/valid.txt -O $WIKI_DATA_DIR/valid.txt.original
wget -x https://dl.fbaipublicfiles.com/colorless-green-rnns/training-data/English/vocab.txt -O $WIKI_DATA_DIR/vocab.txt

Getting the Yedetore CHILDES training data:

• Note: the vocab.txt file provided in pretraining.zip is incorrect. The correct file is at http://www.adityayedetore.com/childes-project/vocab.txt.

CHILDES_DATA_DIR="./data/retraining/childes"
mkdir -p $CHILDES_DATA_DIR
cd /tmp/

wget http://www.adityayedetore.com/childes-project/pretraining.zip
unzip pretraining.zip
cp pretraining/train.txt $CHILDES_DATA_DIR/train.txt.original
cp pretraining/valid.txt $CHILDES_DATA_DIR/valid.txt.original
cp pretraining/test.txt $CHILDES_DATA_DIR/test.txt.original

cd $CHILDES_DATA_DIR
wget http://www.adityayedetore.com/childes-project/vocab.txt # Correct file.

Running retraining for ATB/PG:

PYTHONPATH="." python retraining.py --phenomenon ATB --names tiny --samples_per_condition 100 --seed 1000 --context_size 30 --dataset_name "wikipedia"
PYTHONPATH="." python retraining.py --phenomenon PG --names small --samples_per_condition 100 --seed 1000 --context_size 30 --dataset_name "wikipedia"

Caching surprisal values

Inference is costly, so we cache surprisal values to Redis. If you're running a large scale experiment like the ones in the paper, it's better to run redis before: $ redis-server &.

Modified external sources

The project uses modified versions of the following projects, under modified_external_sources:

lm-povstim-with-childes

Source code for Yedetore et al. (2023).

Changes made

• Changed main.py to support training on original dataset + extra sentences.

colorlessgreenRNNs

Source code for Gulordava et al..

Changes made

• Changed dictionary_corpus.py to support multiple text file sources.