---
license: apache-2.0
language:
- kir
datasets:
- allenai/MADLAD-400
- allenai/nllb
- cis-lmu/Glot500
- oscar-corpus/OSCAR-2109
library_name: transformers
pipeline_tag: text-generation
tags:
- goldfish
- arxiv:2408.10441
---

# kir_cyrl_100mb

Goldfish is a suite of monolingual language models trained for 350 languages.
This model is the <b>Kirghiz</b> (Cyrillic script) model trained on 100MB of data, after accounting for an estimated byte premium of 1.96; content-matched text in Kirghiz takes on average 1.96x 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: kir_cyrl is an [individual language](https://iso639-3.sil.org/code_tables/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)](https://www.arxiv.org/abs/2408.10441).

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

Sample usage also in this Google Colab: [link](https://colab.research.google.com/drive/1rHFpnQsyXJ32ONwCosWZ7frjOYjbGCXG?usp=sharing)

## 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): 196.38MB
* Training text data (byte premium scaled): 100.005MB
* Training tokens: 22320128 (x10 epochs)
* Vocabulary size: 50000
* Compute cost: 1.13932732465152e+17 FLOPs or ~10.8 NVIDIA A6000 GPU hours

Training datasets (percentages prior to deduplication):
* 65.41839%: [MADLAD-400 (CommonCrawl)](https://huggingface.co/datasets/allenai/MADLAD-400)
* 13.62164%: [NLLB (CommonCrawl and ParaCrawl)](https://huggingface.co/datasets/allenai/nllb)
* 9.53594%: [Glot500](https://huggingface.co/datasets/cis-lmu/Glot500), including [CCNet](https://github.com/facebookresearch/cc_net), [Earthlings](https://publicdata.canterbury.ac.nz/Research/Geocorpus/CCGLU_v5.0/), [OSCAR](https://oscar-project.org/), [TIL](https://github.com/turkic-interlingua/til-mt)
* 8.32078%: [OSCAR 2021/09](https://huggingface.co/datasets/oscar-corpus/OSCAR-2109)
* 3.10325%: [Wikipedia 2023/08](https://dumps.wikimedia.org/)


## 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},
}
```