File size: 2,171 Bytes
dacff11
570c514
dacff11
 
570c514
 
 
dacff11
 
29a4019
 
 
 
 
 
364393a
29a4019
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
570c514
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---

language: bn
license: mit
datasets:
- mc4

---

# Bengali GPT-2

Bengali GPT-2 demo. Part of the [Huggingface JAX/Flax event](https://discuss.huggingface.co/t/open-to-the-community-community-week-using-jax-flax-for-nlp-cv/). Also features a [finetuned](https://huggingface.co/khalidsaifullaah/bengali-lyricist-gpt2?) model on bengali song lyrics. 

# Model Description

OpenAI GPT-2 model was proposed in [Language Models are Unsupervised Multitask Learners](https://paperswithcode.com/paper/language-models-are-unsupervised-multitask) paper .Original GPT2 model was a causal (unidirectional) transformer pretrained using language modeling on a very large corpus of ~40 GB of text data. This model has same configuration but has been pretrained on bengali corpus of mC4(multilingual C4) dataset. The code for training the model has all been open-sourced [here](https://huggingface.co/flax-community/gpt2-bengali/tree/main).

# Training Details

Overall Result: 

```Eval loss : 1.45, Eval Perplexity : 3.141```

Data: [mC4-bn](https://huggingface.co/datasets/mc4)

Train Steps: 250k steps

link 🤗 flax-community/gpt2-bengali

Demo : https://huggingface.co/spaces/flax-community/Gpt2-bengali

# Usage 

For using the model there are multiple options available. For example using the pipeline directly we can try to generate sentences.

```
from transformers import pipeline

gpt2_bengali = pipeline('text-generation',model="flax-community/gpt2-bengali", tokenizer='flax-community/gpt2-bengali')
```

Similarly for using the finetuned model on bangla songs we can use following.

```
from transformers import pipeline

singer = pipeline('text-generation',model="khalidsaifullaah/bengali-lyricist-gpt2", tokenizer='khalidsaifullaah/bengali-lyricist-gpt2')
```

For using on other tasks the model needs to be fine-tuned on custom datasets. Details can be found in huggingface [documentation](https://huggingface.co/transformers/training.html)

# Contributors
* Khalid Saifullah
* Tasmiah Tahsin Mayeesha
* Ritobrata Ghosh
* Ibrahim Musa
* M Saiful Bari

### BibTeX entry and citation info

Coming soon! 

<!-- ```bibtex

@inproceedings{...,

  year={2020}

}

``` -->