Modelpublic

goldfish-models/abk_cyrl_full

sourceHugging Faceapache-2.0updated 2y agoView on Hugging Face

1likes19downloads

Model Card

license: apache-2.0 language:

—abk datasets:
—cis-lmu/Glot500
—Nart/abkhaz_text
—legacy-datasets/wikipedia libraryname: transformers pipelinetag: text-generation tags:
—goldfish
—arxiv:2408.10441 ---

abkcyrlfull

Goldfish is a suite of monolingual language models trained for 350 languages. This model is the <b>Abkhazian</b> (Cyrillic script) model trained on 25MB of data (all our data in the language), after accounting for an estimated byte premium of 2.01; content-matched text in Abkhazian takes on average 2.01x as many UTF-8 bytes to encode as English. The Goldfish models are trained primarily for comparability across languages and for low-resource languages; Goldfish performance for high-resource languages is not designed to be comparable with modern large language models (LLMs).

Note: abkcyrl is an [individual language](https://iso639-3.sil.org/codetables/639/data) code. It is not contained in any macrolanguage codes contained in Goldfish (for script cyrl).

All training and hyperparameter details are in our paper, Goldfish: Monolingual Language Models for 350 Languages (Chang et al., 2024).

Training code and sample usage: https://github.com/tylerachang/goldfish

Sample usage also in this Google Colab: link

Model details:

To access all Goldfish model details programmatically, see https://github.com/tylerachang/goldfish/blob/main/model_details.json. All models are trained with a [CLS] (same as [BOS]) token prepended, and a [SEP] (same as [EOS]) token separating sequences. For best results, make sure that [CLS] is prepended to your input sequence (see sample usage linked above)! Details for this model specifically:

—Architecture: gpt2
—Parameters: 124770816
—Maximum sequence length: 512 tokens
—Training text data (raw): 50.86MB
—Training text data (byte premium scaled): 25.245MB
—Training tokens: 6408192 (x10 epochs)
—Vocabulary size: 50000
—Compute cost: 3.2693904211968e+16 FLOPs or ~3.1 NVIDIA A6000 GPU hours

Training datasets (percentages prior to deduplication):

—96.32220%: Glot500, including Nart/abkhaz, Tatoeba, TeDDi, Wikipedia Hugging Face
—3.67618%: Wikipedia 2023/08
—0.00163%: Tatoeba

Citation

If you use this model, please cite:

@article{chang-etal-2024-goldfish,
  title={Goldfish: Monolingual Language Models for 350 Languages},
  author={Chang, Tyler A. and Arnett, Catherine and Tu, Zhuowen and Bergen, Benjamin K.},
  journal={Preprint},
  year={2024},
  url={https://www.arxiv.org/abs/2408.10441},
}