goldfish-models commited on
Commit
58c54be
1 Parent(s): 6a614b7

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +69 -0
README.md ADDED
@@ -0,0 +1,69 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ ---
3
+ license: apache-2.0
4
+ language:
5
+ - tsn
6
+ datasets:
7
+ - allenai/nllb
8
+ - cis-lmu/Glot500
9
+ - sil-ai/bloom-lm
10
+ - statmt/cc100
11
+ - legacy-datasets/wikipedia
12
+ - allenai/MADLAD-400
13
+ library_name: transformers
14
+ pipeline_tag: text-generation
15
+ tags:
16
+ - goldfish
17
+
18
+ ---
19
+
20
+ # tsn_latn_5mb
21
+
22
+ Goldfish is a suite of monolingual language models trained for 350 languages.
23
+ This model is the <b>Tswana</b> (Latin script) model trained on 5MB of data, after accounting for an estimated byte premium of 1.17; content-matched text in Tswana takes on average 1.17x as many UTF-8 bytes to encode as English.
24
+ 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).
25
+
26
+ Note: tsn_latn 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 latn).
27
+
28
+ 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).
29
+
30
+ Training code and sample usage: https://github.com/tylerachang/goldfish
31
+
32
+ Sample usage also in this Google Colab: [link](https://colab.research.google.com/drive/1rHFpnQsyXJ32ONwCosWZ7frjOYjbGCXG?usp=sharing)
33
+
34
+ ## Model details:
35
+
36
+ To access all Goldfish model details programmatically, see https://github.com/tylerachang/goldfish/model_details.json.
37
+ All models are trained with a [CLS] (same as [BOS]) token prepended, and a [SEP] (same as [EOS]) token separating sequences.
38
+ Details for this model specifically:
39
+
40
+ * Architecture: gpt2
41
+ * Parameters: 39087104
42
+ * Maximum sequence length: 512 tokens
43
+ * Training text data (raw): 5.87MB
44
+ * Training text data (byte premium scaled): 5.005MB
45
+ * Training tokens: 1606656 (x10 epochs)
46
+ * Vocabulary size: 50000
47
+ * Compute cost: 1214952599715840.0 FLOPs or ~0.1 NVIDIA A6000 GPU hours
48
+
49
+ Training datasets (percentages prior to deduplication):
50
+ * 54.15070%: [NLLB (CommonCrawl and ParaCrawl)](https://huggingface.co/datasets/allenai/nllb)
51
+ * 25.72794%: [Glot500](https://huggingface.co/datasets/cis-lmu/Glot500), including [AUTSHUMATO](https://autshumato.sourceforge.net/), [BLOOM](https://huggingface.co/datasets/sil-ai/bloom-lm), [CC100](https://huggingface.co/datasets/statmt/cc100), [CORP.NCHLT](https://repo.sadilar.org/handle/20.500.12185/7), [Wortschatz Leipzig Data](https://wortschatz.uni-leipzig.de/en/download), [Mburisano_Covid](https://repo.sadilar.org/handle/20.500.12185/536), [Tatoeba](https://tatoeba.org/en/), [Wikipedia Hugging Face](https://huggingface.co/datasets/legacy-datasets/wikipedia)
52
+ * 18.23011%: [MADLAD-400 (CommonCrawl)](https://huggingface.co/datasets/allenai/MADLAD-400)
53
+ * 1.14046%: [Wikipedia 2023/08](https://dumps.wikimedia.org/)
54
+ * 0.74997%: [eBible](https://ebible.org/find/)
55
+ * 0.00083%: [Tatoeba](https://tatoeba.org/en/)
56
+
57
+
58
+ ## Citation
59
+
60
+ If you use this model, please cite:
61
+
62
+ ```
63
+ @article{chang-etal-2024-goldfish,
64
+ title={Goldfish: Monolingual Language Models for 350 Languages},
65
+ author={Chang, Tyler A. and Arnett, Catherine and Tu, Zhuowen and Bergen, Benjamin K.},
66
+ journal={Preprint},
67
+ year={2024},
68
+ }
69
+ ```