Nonparametric Masked Language Modeling

This repo contains the original implementation of the paper "Nonparametric Masked Language Modeling".

@article{ min2022nonparametric,
    title={ Nonparametric Masked Language Modeling },
    author={ Min, Sewon and Shi, Weijia and Lewis, Mike and Chen, Xilun and Yih, Wen-tau and Hajishirzi, Hannaneh and Zettlemoyer, Luke },
    year={ 2022 }
}

Models are available from Huggingface Hub:hugs:! Check out npm (for phrase retrieval) and npm-single (for token retrieval).

We are working on a simple demo where you can simply download all the resources and deploy on your machine. Stay tuned!

Updates

01/02/2023: The code for training is released. See train.md for instructions.
12/22/2022: The code for inference is released. Stay tuned for the code for training.

Content

Requirements
Download Data
Closed-set Experiments
- Baselines
- NPM
- NPM Single
Open-set Experiments
- Baselines
- NPM
License
Contact

Requirements

conda create -n npm python=3.7
conda activate npm
pip3 install -r requirements.txt --user

If you will use open-set tasks, make sure to install java as well.

conda install -c conda-forge openjdk

Note that multi-gpu inference is not supported for now.

Download Data

Evaluation datasets and reference corpora can be downloaded via

# To run evaluation on closed-set tasks
bash scripts/download_data.sh closed
bash scripts/download_corpus.sh closed

# To run evaluation on open-set tasks
bash scripts/download_data.sh open
bash scripts/download_corpus.sh enwiki

# To run evaluation on TempLAMA (need Wikipedia 2022)
bash scripts/download_data.sh templama
bash scripts/download_corpus.sh new-enwiki

The corpus data is required for NPM and the retrieve-and-generate baselines. If you will only run parametric baselines, you can skip downloading the corpus.

All reference corpus files are saved under corpus/ and evaluation datasets are saved under data/.

Closed-set Experiments

Baselines on closed-set tasks

The following is the script for runing the RoBERTA-large baseline on all 9 datasets used in the paper.

python -m scripts.prompt \
    --checkpoint_path roberta-large \
    --eval_dataset agn+yahoo+rte+subj+sst2+mr+rt+cr+amazon \
    --save_dir save/roberta \
    --single

NPM on closed-set tasks

# To run on AGN, Yahoo and RTE:
bash scripts/save_embeddings.sh npm enwiki-0 false 320
bash scripts/save_embeddings.sh npm cc_news false 320
python -m scripts.prompt \
    --corpus_data enwiki-0+cc_news \
    --checkpoint_path npm \
    --eval_dataset agn+yahoo+rte \
    --temperature 5.0 \
    --save_dir save/npm

# To run on Subj:
bash scripts/save_embeddings.sh npm subj false 320
python -m scripts.prompt \
    --corpus_data subj \
    --checkpoint_path npm \
    --eval_dataset subj \
    --temperature 5.0 \
    --save_dir save/npm

# To run on SST-2, MR, RT, CR and Amazon:
bash scripts/save_embeddings.sh npm imdb false 320
bash scripts/save_embeddings.sh npm amazon false 320
python -m scripts.prompt \
    --corpus_data imdb+amazon \
    --checkpoint_path npm \
    --eval_dataset sst2+mr+rt+cr+amazon \
    --temperature 5.0 \
    --save_dir save/npm

Note that scripts/save_embeddings.sh takes

model name (npm or npm-single)
corpus name
whether it is an open-set task (true or false)
batch size (320 is good for a 32gb GPU; if trainer.precision=16 is used, 400 is good for a 32gb GPU) as arguments. Embeddings are saved under save/{model_name}/dstore.

NPM Single on closed-set tasks

# To run on AGN, Yahoo and RTE:
bash scripts/save_embeddings.sh npm-single enwiki-0 false 320
bash scripts/save_embeddings.sh npm-single cc_news false 320
python -m scripts.prompt \
    --corpus_data enwiki-0+cc_news \
    --checkpoint_path npm-single \
    --eval_dataset agn+yahoo+rte \
    --temperature 5.0 \
    --single \
    --save_dir save/npm-single

# To run on Subj:
bash scripts/save_embeddings.sh npm-single subj false 320
python -m scripts.prompt \
    --corpus_data subj \
    --checkpoint_path npm-single \
    --eval_dataset subj \
    --temperature 5.0 \
    --single \
    --save_dir save/npm-single

# To run on SST-2, MR, RT, CR and Amazon:
bash scripts/save_embeddings.sh npm-single imdb false 320
bash scripts/save_embeddings.sh npm-single amazon false 320
python -m scripts.prompt \
    --corpus_data imdb+amazon \
    --checkpoint_path npm-single \
    --eval_dataset sst2+mr+rt+cr+amazon \
    --temperature 5.0 \
    --single \
    --save_dir save/npm-single

Open-set Experiments

Baselines on open-set tasks

Run the following to run causal language model baselines (T5 baselines are TBA!).

python -m scripts.clm_prompt \
    --eval_dataset {lama-trex|lama-google_re|kamel|triviaqa|nq|entity_translation} \
    --model_name {j-6b|neo-1.3b|neo-2.7b|neox-20b|opt-1.3b|opt-2.7b|opt-6.7b|opt-13b|opt-30b|bloom-1b7|bloom-3b|bloom-7b1} \
    --save_dir save

By default, this does not use any passages from an external corpus. Specify --ret bm25 if use BM25 passages from Wikipedia 2019, and --ret bm25_2022 to use BM25 passages from Wikipedia 2022 (for TempLAMA).

NPM on open-set tasks

Please note that running open-set tasks requires around 70GB of RAM and 1.4TB of disk memory. If you want to reduce the RAM usage, you can specify --keep_uint8 while running python -m scripts.prompt below, which reduces the RAM usage from 70GB to 40GB while increasing the datastore setting time. We will explore further optimizing RAM/disk usage in the future version of the code (PR is also welcome!).

# Note that this can be executed in parallel with up to 20 GPUs. In total, it takes about 10 GPU hours and 1.4TB of disk memory.
for i in {0..19} ; do
    bash scripts/save_embeddings.sh npm enwiki-${i} true 320
done

# Loading the model takes about 40min, and 70GB of RAM (specify `--keep_uint8` to reduce RAM usage to 40GB which increases the model loading time to 60-80min).
python -m scripts.prompt \
    --corpus_data enwiki \
    --checkpoint_path npm \
    --eval_dataset lama-trex+lama-google_re+kamel+triviaqa+nq+entity_translation \
    --save_dir save/npm \
    --remove_stopwords \
    --restricted \
    --open

To evaluate on TempLAMA, use new-enwiki instead of enwiki, and use --eval_dataset {templama|unchanged_templama}.

License

NPM is CC-BY-NC 4.0 licensed.

Contact

Please leave Github issues or contact Sewon Min sewon@cs.washington.edu for any questions.

facebookresearch/NPM