Trained on 100 million words and still in shape:

BERT meets British National Corpus

David Samuel, Andrey Kutuzov, Lilja Øvrelid and Erik Velldal

University of Oslo
Language Technology Group

Paper
HuggingFace models

---

This is the official repository for our EACL paper about pre-training language models on a representative 100M-word corpus. We propose a data-efficient LM architecture (LTG-BERT) that outperforms the original BERT model. We believe that this type of modestly-sized, but representative, corpora has great potential as a language modeling benchmark.

---

Content of this repository

• ./modeling_ltgbert.py: HuggingFace-compatible wrapper for LTG-BERT
• ./preprocessing/: Scripts for processing the XML format of BNC and for processing the evaluation datasets
• ./training/: Scripts for training LTG-BERT on processed BNC
• ./evaluation/: Evaluation scripts for evaluation LTG-BERT on (Super)GLUE, edge probing and BLiMP

---

Please cite the following publication (just arXiv for now)

@inproceedings{samuel-etal-2023-trained,
    title = "Trained on 100 million words and still in shape: {BERT} meets {B}ritish {N}ational {C}orpus",
    author = "Samuel, David  and
      Kutuzov, Andrey  and
      {\O}vrelid, Lilja  and
      Velldal, Erik",
    booktitle = "Findings of the Association for Computational Linguistics: EACL 2023",
    month = may,
    year = "2023",
    address = "Dubrovnik, Croatia",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.findings-eacl.146",
    pages = "1954--1974",
    abstract = "While modern masked language models (LMs) are trained on ever larger corpora, we here explore the effects of down-scaling training to a modestly-sized but representative, well-balanced, and publicly available English text source {--} the British National Corpus. We show that pre-training on this carefully curated corpus can reach better performance than the original BERT model. We argue that this type of corpora has great potential as a language modeling benchmark. To showcase this potential, we present fair, reproducible and data-efficient comparative studies of LMs, in which we evaluate several training objectives and model architectures and replicate previous empirical results in a systematic way. We propose an optimized LM architecture called LTG-BERT.",
}