File size: 2,889 Bytes
fb8e90c
 
426172e
 
 
 
 
 
fb8e90c
426172e
 
 
 
 
8d066de
426172e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
73
74
75
76
77
78
79
80
81
82
83
---
license: apache-2.0
language: 
  - ind
datasets:
  - uonlp/CulturaX
tags:
  - t5
---

## IndoNanoT5 Base

IndoNanoT5 Base is an Indonesian sequence-to-sequence language model based on the [T5](https://arxiv.org/abs/1910.10683) architecture. We conducted pre-training on an open-source Indonesian corpus of [uonlp/CulturaX](https://huggingface.co/datasets/uonlp/CulturaX). On a held-out subset of the corpus, our model achieved an evaluation loss of 2.082 or a perplexity of about 8.02.

This model was trained using the [nanoT5](https://github.com/PiotrNawrot/nanoT5) PyTorch framework. All training was done on an NVIDIA H100 GPU. [LazarusNLP/IndoNanoT5-base](https://huggingface.co/LazarusNLP/IndoNanoT5-base) is released under Apache 2.0 license.

## Model Detail

- **Developed by**: [LazarusNLP](https://lazarusnlp.github.io/)
- **Model type**: Encoder-decoder T5 transformer language model
- **Language(s)**: Indonesian
- **License**: [Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0.html)
- **Contact**: [Wilson Wongso](https://wilsonwongso.dev/)

## Use in 🤗Transformers

```python
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

model_checkpoint = "LazarusNLP/IndoNanoT5-base"

tokenizer = AutoTokenizer.from_pretrained(model_checkpoint)
model = AutoModelForSeq2SeqLM.from_pretrained(model_checkpoint)
```

## Training Datasets

Around 4B tokens from the following corpora were used during pre-training.

- [Cleaned, Enormous, and Public: The Multilingual Fuel to Democratize Large Language Models for 167 Languages](https://huggingface.co/datasets/uonlp/CulturaX)

## Training Hyperparameters

The following hyperparameters were used during training:

- `total_steps`: 65536
- `input_length`: 512
- `batch_size`: 128
- `grad_acc`: 1
- `base_lr`: 5e-3
- `optimizer`: AdamWScaled with `betas=(0.9,0.999)` and `epsilon=1e-08`
- `weight_decay`: 0.0
- `lr_scheduler`: cosine
- `warmup_steps`: 10000
- `final_cosine`: 1e-5
- `grad_clip`: 1.0
- `precision`: `bf16`

## Acknowledgements

We would like to acknowledge [nanoT5](https://github.com/PiotrNawrot/nanoT5) for inspiring this project.

## Credits

BhinnekaLM is developed with love by:

<div style="display: flex;">
<a href="https://github.com/anantoj">
    <img src="https://github.com/anantoj.png" alt="GitHub Profile" style="border-radius: 50%;width: 64px;margin:0 4px;">
</a>

<a href="https://github.com/DavidSamuell">
    <img src="https://github.com/DavidSamuell.png" alt="GitHub Profile" style="border-radius: 50%;width: 64px;margin:0 4px;">
</a>

<a href="https://github.com/stevenlimcorn">
    <img src="https://github.com/stevenlimcorn.png" alt="GitHub Profile" style="border-radius: 50%;width: 64px;margin:0 4px;">
</a>

<a href="https://github.com/w11wo">
    <img src="https://github.com/w11wo.png" alt="GitHub Profile" style="border-radius: 50%;width: 64px;margin:0 4px;">
</a>
</div>