Skip to content

citation_guide.md

Latest commit

 

History

History
126 lines (93 loc) · 4.66 KB

citation_guide.md

File metadata and controls

126 lines (93 loc) · 4.66 KB
Raw

Citation guide

This page is for researchers who want to cite, compare, or reproduce NGramStore (formerly me_fasttext).

For a one-page summary of the artifact, including fit, claims, evidence, and non-goals, see artifact_card.md.

Cite the paper

If the method, implementation, benchmark protocol, or compact serving layout is useful in your work, please cite the arXiv paper:

@misc{du2026memoryefficientfasttext,
  title         = {Memory-Efficient FastText: A Comprehensive Approach Using Double-Array Trie Structures and Mark-Compact Memory Management},
  author        = {Yimin Du},
  year          = {2026},
  eprint        = {2506.01254},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2506.01254}
}

Cite the native n-gram artifact

NGramStore is the renamed repository, with the original paper title and method attribution retained. The C gather ABI, immutable PyTorch adapters and Engram-related experiments are later software work. Cite their exact repository commit/report alongside the FastText paper when relevant; do not attribute the new results to that paper. Replace the bracketed commit below with the full SHA used in your experiment:

@misc{du2026mefasttextngramstorage,
  author       = {Yimin Du},
  title        = {NGramStore: Compression, Storage and Exact Lookup for Learned N-Gram Vectors},
  year         = {2026},
  howpublished = {Software and reproducibility artifact},
  url          = {https://github.com/initial-d/ngram-store},
  note         = {Commit [full SHA]; report [report path and date]}
}

This is a repository/software citation, not a new publication or minted DOI. The Engram quickstart connects code, source-pinned verification, trained results and comparison requirements.

Suggested wording for the tested boundary:

The NGramStore engine provides immutable storage/gather adapters that preserve original hash IDs, head offsets and row bytes. Small source-pinned Engram demo checks also preserve gated module output. Separate trained TinyEngram/Qwen CPU/CUDA checks and timing controls do not establish official trained DeepSeek Engram or production serving results.

For approximate compression, name the actual trained table/corpus, threshold, norm guard, byte denominator and distortion metric. Do not describe the character-ngram reference's compression factor as an Engram model result.

When to cite the method

me_fasttext is most relevant when your work discusses:

  • FastText-style word or subword embeddings
  • OOV-heavy retrieval, ranking, or classification
  • hash collisions in character n-gram embedding tables
  • trie-backed dictionaries or double-array trie lookup for NLP
  • memory-efficient embedding storage
  • mmap-based model serving
  • compact inference artifacts for large vocabularies
  • lexical retrieval layers used before dense reranking

Short related-work wording

Use or adapt one of these sentences when citing the paper.

For FastText and subword embeddings:

Du revisits FastText-style subword embeddings from a systems perspective, replacing hash-bucketed n-gram ids with exact trie-backed ids before compact serving export.

For memory-efficient serving:

me_fasttext separates n-gram identity from serving layout: it trains with exact subword ids, then rewrites retained rows into a compact mmap-ready artifact.

For retrieval pipelines:

The method is useful as a low-cost lexical embedding layer for OOV-heavy retrieval and candidate generation before dense reranking.

For data-structure readers:

The implementation uses trie-backed word and n-gram dictionaries so lexical identities remain inspectable before conservative row sharing.

Reproduction checklist

When comparing against this work, please report:

  • git commit and release tag
  • corpus language, token count, document count, and preprocessing
  • vocabulary size and extracted n-gram count
  • training mode and hyperparameters
  • memory, artifact size, cold load time, and warm query latency
  • downstream quality metric
  • hardware, storage, OS, compiler, and filesystem

The benchmark protocol includes a fuller report template. The citation evidence ladder explains which reports are strong enough to link from README-level docs or reuse as paper/report evidence.

Code citation metadata

This repository includes:

  • CITATION.cff for GitHub citation UI and reference managers
  • .zenodo.json for DOI-ready Zenodo metadata
  • release notes for the v0.1.0 research prototype

If a Zenodo DOI is enabled later, cite the DOI for the exact software snapshot and cite the arXiv paper for the method.