goldfish-models
commited on
Commit
•
ce7a823
1
Parent(s):
26956dd
Upload README.md with huggingface_hub
Browse files
README.md
ADDED
@@ -0,0 +1,65 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
|
2 |
+
---
|
3 |
+
license: apache-2.0
|
4 |
+
language:
|
5 |
+
- bua
|
6 |
+
datasets:
|
7 |
+
- allenai/MADLAD-400
|
8 |
+
- oscar-corpus/OSCAR-2109
|
9 |
+
library_name: transformers
|
10 |
+
pipeline_tag: text-generation
|
11 |
+
tags:
|
12 |
+
- goldfish
|
13 |
+
|
14 |
+
---
|
15 |
+
|
16 |
+
# bua_cyrl_10mb
|
17 |
+
|
18 |
+
Goldfish is a suite of monolingual language models trained for 350 languages.
|
19 |
+
This model is the <b>Buriat</b> (Cyrillic script) model trained on 10MB of data, after accounting for an estimated byte premium of 1.70; content-matched text in Buriat takes on average 1.70x as many UTF-8 bytes to encode as English.
|
20 |
+
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).
|
21 |
+
|
22 |
+
Note: bua_cyrl is a [macrolanguage](https://iso639-3.sil.org/code_tables/639/data) code. Individual language code bxr_cyrl (Russia Buriat) is included in Goldfish, although with less data.
|
23 |
+
|
24 |
+
All training and hyperparameter details are in our paper, [Goldfish: Monolingual Language Models for 350 Languages (Chang et al., 2024)](https://github.com/tylerachang/goldfish/blob/main/goldfish_paper_20240815.pdf).
|
25 |
+
|
26 |
+
Training code and sample usage: https://github.com/tylerachang/goldfish
|
27 |
+
|
28 |
+
Sample usage also in this Google Colab: [link](https://colab.research.google.com/drive/1rHFpnQsyXJ32ONwCosWZ7frjOYjbGCXG?usp=sharing)
|
29 |
+
|
30 |
+
## Model details:
|
31 |
+
|
32 |
+
To access all Goldfish model details programmatically, see https://github.com/tylerachang/goldfish/model_details.json.
|
33 |
+
All models are trained with a [CLS] (same as [BOS]) token prepended, and a [SEP] (same as [EOS]) token separating sequences.
|
34 |
+
Details for this model specifically:
|
35 |
+
|
36 |
+
* Architecture: gpt2
|
37 |
+
* Parameters: 39087104
|
38 |
+
* Maximum sequence length: 512 tokens
|
39 |
+
* Training text data (raw): 17.02MB
|
40 |
+
* Training text data (byte premium scaled): 10.005MB
|
41 |
+
* Training tokens: 2348544 (x10 epochs)
|
42 |
+
* Vocabulary size: 50000
|
43 |
+
* Compute cost: 1775011579822080.0 FLOPs or ~0.2 NVIDIA A6000 GPU hours
|
44 |
+
|
45 |
+
Training datasets (percentages prior to deduplication):
|
46 |
+
* 89.97116%: [Languages of Russia](http://web-corpora.net/wsgi3/minorlangs/download)
|
47 |
+
* 8.41374%: [MADLAD-400 (CommonCrawl)](https://huggingface.co/datasets/allenai/MADLAD-400)
|
48 |
+
* 1.21634%: [Wikipedia 2023/08](https://dumps.wikimedia.org/)
|
49 |
+
* 0.38769%: [Wortschatz Leipzig Data](https://wortschatz.uni-leipzig.de/en/download)
|
50 |
+
* 0.00722%: [Tatoeba](https://tatoeba.org/en/)
|
51 |
+
* 0.00386%: [OSCAR 2021/09](https://huggingface.co/datasets/oscar-corpus/OSCAR-2109)
|
52 |
+
|
53 |
+
|
54 |
+
## Citation
|
55 |
+
|
56 |
+
If you use this model, please cite:
|
57 |
+
|
58 |
+
```
|
59 |
+
@article{chang-etal-2024-goldfish,
|
60 |
+
title={Goldfish: Monolingual Language Models for 350 Languages},
|
61 |
+
author={Chang, Tyler A. and Arnett, Catherine and Tu, Zhuowen and Bergen, Benjamin K.},
|
62 |
+
journal={Preprint},
|
63 |
+
year={2024},
|
64 |
+
}
|
65 |
+
```
|