| --- |
| license: other |
| language: |
| - en |
| pretty_name: MoNIM |
| tags: |
| - continual-learning |
| - language-modeling |
| - knn-lm |
| - semiparametric-models |
| - news-crawl |
| - monim |
| - memory |
| --- |
| |
| # MoNIM |
|
|
| This repository contains the days 1-30 daily data and the GPT-2 small checkpoint used for reproducing the continual-learning experiments in the ACL 2025 paper [Learn to Memorize: Scalable Continual Learning in Semiparametric Models with Mixture-of-Neighbors Induction Memory](https://aclanthology.org/2025.acl-long.1385/). |
|
|
| ## Dataset Summary |
|
|
| The paper uses News Crawl data from the first half of 2020 (NC-20H1) as streaming data for continual learning. This repository provides the first 30 daily shards from the prepared `daily` data, intended for reproduction research. |
|
|
| Each daily shard includes: |
|
|
| - `train.txt`, `valid.txt`, `test.txt`: article-level text splits. |
| - `train.bpe`, `valid.bpe`, `test.bpe`: GPT-2 BPE-tokenized text. |
| - `bin/`: fairseq mmap binaries built from the BPE files. |
| - `net/train.*`, `net/test.*`: small held-out subsets used by the MoNIM auxiliary selection network. |
| - `net/total/`: combined auxiliary-network inputs and their fairseq binaries. |
| - `freq_cache_id.pickle`, `fertility_cache_id.pickle`: cached token statistics used by the reproduction code. |
|
|
| Top-level tokenizer files: |
|
|
| - `dict.txt` |
| - `encoder.json` |
| - `vocab.bpe` |
|
|
| Model checkpoint: |
|
|
| - `checkpoints/gpt2-small/checkpoint_best.pt` |
|
|
| KNN datastore/index artifacts, MoNIM adapter checkpoints, run logs, watcher scripts, and experiment outputs are not part of this repository. |
|
|
| ## Data Source |
|
|
| The source corpus is WMT News Crawl: |
|
|
| - News Crawl index: https://data.statmt.org/news-crawl/ |
| - News Crawl README: https://data.statmt.org/news-crawl/README |
|
|
| The News Crawl README describes the corpus as monolingual text extracted from online newspapers and released for WMT shared tasks. It also states that News Crawl does not own the extracted text and that only the packaging is released under Creative Commons CC0. Users of this repository must comply with the original News Crawl terms, the upstream take-down policy, and any applicable rights of the original text sources. |
|
|
| ## Preparation |
|
|
| The prepared daily data follows the paper setup: approximately 98% of articles are used for training, 1% for validation, and 1% for test. Text is tokenized with a GPT-2 BPE tokenizer, then preprocessed into fairseq mmap binaries with the shared dictionary. |
|
|
| ## Intended Use |
|
|
| This repository is intended for academic research and reproduction of the MoNIM continual-learning experiments. It is not intended as a general-purpose news corpus release. |
|
|
| ## Citation |
|
|
| If you use this dataset, cite the MoNIM paper and the upstream News Crawl/WMT data source: |
|
|
| ```bibtex |
| @inproceedings{peng-etal-2025-learn, |
| title = "Learn to Memorize: Scalable Continual Learning in Semiparametric Models with Mixture-of-Neighbors Induction Memory", |
| author = "Peng, Guangyue and |
| Ge, Tao and |
| Luo, Wen and |
| Li, Wei and |
| Wang, Houfeng", |
| booktitle = "Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)", |
| month = jul, |
| year = "2025", |
| address = "Vienna, Austria", |
| publisher = "Association for Computational Linguistics", |
| url = "https://aclanthology.org/2025.acl-long.1385/", |
| doi = "10.18653/v1/2025.acl-long.1385", |
| pages = "28517--28531", |
| } |
| ``` |
|
|
