File size: 3,606 Bytes
8c6d716
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
dc4fbbe
8c6d716
 
 
 
 
 
 
 
 
 
dc4fbbe
8c6d716
 
 
 
 
 
 
71a0cb8
8c6d716
dc4fbbe
8c6d716
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
dc4fbbe
8c6d716
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72

---
license: apache-2.0
language:
- tir
datasets:
- allenai/nllb
- allenai/MADLAD-400
- cis-lmu/Glot500
- castorini/afriberta-corpus
- legacy-datasets/wikipedia
- csebuetnlp/xlsum
library_name: transformers
pipeline_tag: text-generation
tags:
- goldfish
- arxiv:2408.10441
---

# tir_ethi_full

Goldfish is a suite of monolingual language models trained for 350 languages.
This model is the <b>Tigrinya</b> (Ge'ez script) model trained on 252MB of data (all our data in the language), after accounting for an estimated byte premium of 1.76; content-matched text in Tigrinya takes on average 1.76x 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: tir_ethi 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 ethi).

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): 446.00MB
* Training text data (byte premium scaled): 252.985MB
* Training tokens: 56515072 (x10 epochs)
* Vocabulary size: 50000
* Compute cost: 2.88450203222016e+17 FLOPs or ~27.3 NVIDIA A6000 GPU hours

Training datasets (percentages prior to deduplication):
* 50.52196%: [Tigrinya Language Modeling Dataset](https://zenodo.org/record/5139094)
* 20.99445%: [NLLB (CommonCrawl and ParaCrawl)](https://huggingface.co/datasets/allenai/nllb)
* 13.46805%: [MADLAD-400 (CommonCrawl)](https://huggingface.co/datasets/allenai/MADLAD-400)
* 10.08565%: [Glot500](https://huggingface.co/datasets/cis-lmu/Glot500), including [AfriBERTa](https://huggingface.co/datasets/castorini/afriberta-corpus), [HornMT](https://github.com/asmelashteka/HornMT), [Wortschatz Leipzig Data](https://wortschatz.uni-leipzig.de/en/download), [MoT](https://github.com/bltlab/mot), [Parallel Corpora for Ethiopian Languages](https://github.com/AAUThematic4LT/Parallel-Corpora-for-Ethiopian-Languages), [TICO](https://tico-19.github.io/), [Wikipedia Hugging Face](https://huggingface.co/datasets/legacy-datasets/wikipedia), [XLSum](https://huggingface.co/datasets/csebuetnlp/xlsum)
* 4.86709%: [AfriBERTa](https://huggingface.co/datasets/castorini/afriberta-corpus)
* 0.06280%: [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},
}
```