NPM-single

NPM-single is a nonparametric masked language model, pretrained on English text data. It was introduced by "Nonparametric Masked Language Modeling" and first released in facebookresearch/NPM .

Model description

NPM consists of an encoder and a reference corpus, and models a nonparametric distribution over a reference corpus. The key idea is to map all the phrases in the corpus into a dense vector space using the encoder and, when given a query with a MASK at inference, use the encoder to locate the nearest phrase from the corpus and fill in the MASK.

NPM-single is a variant of NPM that retrieves a token from the corpus, instead of a phrase.

Intended uses & limitations

While this repo includes the encoder weights, NPM-single has to be used together with a datstore. For more details on how to use NPM-single, please refer to the original repo .

Note that this model is primarily for filling in a MASK token. Future work can investigate how to use NPM-single for text generation.

Training procedure

NPM-single was trained on English Wikipedia (August 2019) and an English portion of CC-News (Mackenzie et al. (2020), February 2019), which contains 13B tokens in total. NPM-single used the model architecture and initial weights of RoBERTa large (Liu et al., 2019), consisting of 354M parameters. Training is done for 100,000 steps, using thirty-two 32GB GPUs.

More details about training can be found in the paper . Code for training NPM-single can be found in the original repo .

Evaluation results

NPM-single is evaluated on nine closed-set tasks (tasks with a small set of options given). NPM-single consistently outperforms significantly larger models such as GPT-3 and T5. Detailed results can be found from the paper .

BibTeX entry and citation info

@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 }
}