Commit
•
e60f2d6
1
Parent(s):
34f3064
Update README.md
Browse files
README.md
CHANGED
@@ -1,19 +1,20 @@
|
|
1 |
|
2 |
# text_generation_bangla_model
|
3 |
BanglaCLM dataset:
|
4 |
-
OSCAR: 12.84GB
|
5 |
-
Wikipedia dump: 6.24GB
|
6 |
-
ProthomAlo: 3.92GB
|
7 |
-
Kalerkantho: 3.24GB
|
8 |
|
|
|
9 |
|
10 |
-
|
|
|
|
|
|
|
|
|
11 |
|
12 |
-
More information needed
|
13 |
|
14 |
-
##
|
|
|
|
|
15 |
|
16 |
-
More information needed
|
17 |
|
18 |
## Training and evaluation data
|
19 |
The BanglaCLM data set is divided into a training set (90%)and a validation set (10%).
|
@@ -24,18 +25,30 @@ The BanglaCLM data set is divided into a training set (90%)and a validation set
|
|
24 |
### Training hyperparameters
|
25 |
|
26 |
The following hyperparameters were used during training:
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
|
33 |
-
|
34 |
-
|
35 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
36 |
- training_precision: float32
|
37 |
|
|
|
38 |
### Training results
|
|
|
39 |
perplexity score: 2.86.
|
40 |
|
41 |
|
@@ -45,3 +58,18 @@ perplexity score: 2.86.
|
|
45 |
- TensorFlow 2.11.0
|
46 |
- Datasets 2.10.0
|
47 |
- Tokenizers 0.13.2
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
|
2 |
# text_generation_bangla_model
|
3 |
BanglaCLM dataset:
|
|
|
|
|
|
|
|
|
4 |
|
5 |
+
- OSCAR: 12.84GB
|
6 |
|
7 |
+
- Wikipedia dump: 6.24GB
|
8 |
+
|
9 |
+
- ProthomAlo: 3.92GB
|
10 |
+
|
11 |
+
- Kalerkantho: 3.24GB
|
12 |
|
|
|
13 |
|
14 |
+
## Model description
|
15 |
+
|
16 |
+
- context size : 128
|
17 |
|
|
|
18 |
|
19 |
## Training and evaluation data
|
20 |
The BanglaCLM data set is divided into a training set (90%)and a validation set (10%).
|
|
|
25 |
### Training hyperparameters
|
26 |
|
27 |
The following hyperparameters were used during training:
|
28 |
+
|
29 |
+
- Batch size: 32
|
30 |
+
|
31 |
+
- Initial learning rate: 5e-5
|
32 |
+
|
33 |
+
- Number of warmup steps: 10000
|
34 |
+
|
35 |
+
- Weight decay rate: 0.01
|
36 |
+
|
37 |
+
- Tokenization algorithm: BPE
|
38 |
+
|
39 |
+
- Vocabulary size of tokenizer: 50256
|
40 |
+
|
41 |
+
- Total trainable params: 124,439,808
|
42 |
+
|
43 |
+
- Epochs: 40
|
44 |
+
|
45 |
+
- Number of training steps: 40772228
|
46 |
+
|
47 |
- training_precision: float32
|
48 |
|
49 |
+
|
50 |
### Training results
|
51 |
+
|
52 |
perplexity score: 2.86.
|
53 |
|
54 |
|
|
|
58 |
- TensorFlow 2.11.0
|
59 |
- Datasets 2.10.0
|
60 |
- Tokenizers 0.13.2
|
61 |
+
|
62 |
+
### Citation
|
63 |
+
If you find this model helpful, please cite.
|
64 |
+
```
|
65 |
+
@INPROCEEDINGS{10303383,
|
66 |
+
author={Salim, Md. Shahidul and Murad, Hasan and Das, Dola and Ahmed, Faisal},
|
67 |
+
booktitle={2023 International Conference on Information and Communication Technology for Sustainable Development (ICICT4SD)},
|
68 |
+
title={BanglaGPT: A Generative Pretrained Transformer-Based Model for Bangla Language},
|
69 |
+
year={2023},
|
70 |
+
volume={},
|
71 |
+
number={},
|
72 |
+
pages={56-59},
|
73 |
+
doi={10.1109/ICICT4SD59951.2023.10303383}}
|
74 |
+
|
75 |
+
```
|