---
language:
- en
- hi
---
# Reproducing GPT-2 for Indic Languages

This Dataset is part of my project on "Reproducing GPT-2 for Indic Languages". Check out the main repository: [https://github.com/Shaligram-Dewangan/GPT-2-for-Indic-Languages](https://github.com/Shaligram-Dewangan/GPT-2-for-Indic-Languages)

## Dataset Description

This is a tokenized text data for pre-training small sized language models and learning purposes.

This dataset contains approximately 20 billion English + Hindi language tokens. It is tokenized using a [custom BPE Tokeinzer](https://github.com/Shaligram-Dewangan/GPT-2-for-Indic-Languages/blob/main/tokenizer/tokenizer.json). The English dataset is from [Fineweb-Edu](https://huggingface.co/spaces/HuggingFaceFW/blogpost-fineweb-v1) [sample-10BT](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu/tree/main/sample/10BT) and Hindi dataset is from [Fineweb-2](https://huggingface.co/spaces/HuggingFaceFW/blogpost-fine-tasks) [hin_Deva](https://huggingface.co/datasets/HuggingFaceFW/fineweb-2/tree/main/data/hin_Deva/train) dataset.

## Dataset Card

|   **Attribute**    |        **Details**        | 
|:-------------------|:-------------------------:| 
| Total size         | ~20.6 Billion tokens      | 
| Size on disk       | ~40 GB                    | 
| Languages          | English and Hindi         |
| English tokens     | ~10.6 Billion             | 
| Hindi tokens       | ~10 Billion               | 
| Mixing type        | [Random mixing](https://github.com/Shaligram-Dewangan/GPT-2-for-Indic-Languages/blob/main/data/tokenize_datasets.py) | 
| Tokenizer          | [BPE tokenizer](https://github.com/Shaligram-Dewangan/GPT-2-for-Indic-Languages/blob/main/tokenizer/tokenizer.json)  | 
| Data Source        | [Fineweb-v1](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu) and [Fineweb-v2](https://huggingface.co/datasets/HuggingFaceFW/fineweb-2) |

## References

1. Penedo et al. *The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale*, 2024. [Blog](https://huggingface.co/spaces/HuggingFaceFW/blogpost-fineweb-v1) · [Fineweb-Edu Repo](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu)

2. Kydlicek et al. *FineTasks: Finding signal in a haystack of 200+ multilingual tasks*, 2024. [Blog](https://huggingface.co/spaces/HuggingFaceFW/blogpost-fine-tasks) · [Fineweb-2 Repo](https://huggingface.co/datasets/HuggingFaceFW/fineweb-2)