|
--- |
|
license: apache-2.0 |
|
widget: |
|
- text: "अपने अनुप्रयोग को पहुंचनीयता व्यायाम" |
|
- text: "जनतंत्र की सफलता केवल इस बात से नहीं हो सकती है कि हर" |
|
- text: "अगर इसके बाद भी वे फैसले पर कायम रहते हैं और" |
|
- text: "मामले का खुलासा होने के बाद" |
|
- text: "My name is Julien and I like to" |
|
- text: "My name is Thomas and my main" |
|
inference: |
|
parameters: |
|
max_length: 200 |
|
--- |
|
|
|
# Model Overview: |
|
The model is a language generation model designed for extending the GPT2 models to support Hindi language along with the original languages that it supports. It was fine-tuned on Hindi texts of [wikipedia](https://www.kaggle.com/datasets/disisbig/hindi-wikipedia-articles-55k) articles. |
|
|
|
# Model Architecture and Parameters: |
|
The model architecture is based on the GPT-2 framework, specifically using the parameters of the small version of the original OpenAI GPT2 model. It employs a Byte Pair Encoding (BPE) tokenizer. |
|
|
|
# Corpus: |
|
The training corpus for Hindi GPT2 consists of Wikipedia articles. |
|
|
|
# Tokenizer: |
|
A tokenizer is trained on Hindi Wikipedia Corpus. The new tokenizer vocabulary (5000 tokens) is merged with existing tokenizer. Hindi GPT2 uses a byte-level version of Byte Pair Encoding (BPE) for tokenizing Hindi text, including Unicode characters. The tokenizer has a vocabulary size of 53497, which allows it to effectively represent the Hindi language's rich vocabulary. Input sequences are formed by breaking the text into consecutive tokens with a maximum length of 1024 tokens. |
|
|
|
## Intended uses & limitations |
|
|
|
More information needed |
|
|
|
## Training and evaluation data |
|
|
|
More information needed |
|
|
|
## Training procedure |
|
|
|
More information needed |
|
|
|
### Training hyperparameters |
|
|
|
The following hyperparameters were used during training: |
|
- learning_rate: 0.0005 |
|
- train_batch_size: 64 |
|
- eval_batch_size: 64 |
|
- seed: 42 |
|
- gradient_accumulation_steps: 4 |
|
- total_train_batch_size: 256 |
|
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 |
|
- lr_scheduler_type: cosine |
|
- lr_scheduler_warmup_steps: 500 |
|
- num_epochs: 1 |
|
- mixed_precision_training: Native AMP |
|
|
|
### Training results |
|
|
|
| Step | Training Loss | Validation Loss | |
|
| :---- | :------------- | :--------------- | |
|
| 500 | 2.0016 | 1.066703 | |
|
| 1000 | 1.0314 | 0.959653 | |
|
| 1500 | 0.9593 | 0.918827 | |
|
| 2000 | 0.922 | 0.889607 | |
|
| 2500 | 0.8983 | 0.872523 | |
|
| 3000 | 0.8852 | 0.863592 | |
|
|
|
|
|
### Framework versions |
|
|
|
- Transformers 4.30.2 |
|
- torch 1.13.1 |
|
- Datasets 2.13.1 |
|
- Tokenizers 0.13.3 |
|
|
