README.md

---

language: ta
license: MIT
datasets:
- OSCAR
- IndicNLP
widget:
- text: 'ஒரு ஊரிலே ஒரு காக்கைக்கு'

---
# GPT2-Tamil

This repository is created as part of the Flax/Jax community week by Huggingface. The aim of this project is to train a language model using GPT-2 specifically for Tamil language.

## Setup:
To setup the project, run the following command,
``` pip install -r requirements.txt
```
 
## Dataset Used:
The GTP-2 model is trained using OSCAR (Tamil) and IndicNLP (Tamil) dataset

## Train model:
To perform training, do the following steps,

- Export the model directory (where you want to store the model artifacts like config, tokenizer, etc.)
```
 export MODEL_DIR=<model_dir>
```
- Create the config.json by running the following command,
```
 python src/create_config.py
```
- Create the tokenizer by running the following command,
```
 python src/train_tokenizer.py
```
- Once the config and tokenizer is created, run the following script to start training the flax model
```
 python scripts/train_gpt2-oscar-tamil.sh
```

## Inference:
To perform language generation using the model, pipeline can be used directly.

- First convert the flax model to pytorch using the following command,
```
 python src/convert_flax_to_pytorch.py
```
- Use the following snippet to perform language generation,
```
 from transformers import AutoTokenizer, AutoModelWithLMHead, pipeline
 model_name = 'abinayam/gpt-2-tamil'
 model = AutoModelWithLMHead.from_pretrained(model_name)
 tokenizer = AutoTokenizer.from_pretrained(model_name)
 input_text = "ஒரு ஊரிலே ஒரு காக்கைக்கு"
 max_len = 300
 generator = pipeline('text-generation', model=model, tokenizer=tokenizer)
 sequence = generator(input_text, max_length=max_len)
```