hishab/titulm-bangla-corpus

TituLM Bangla Corpus

This dataset is associated with the paper TituLLMs: A Family of Bangla LLMs with Comprehensive Benchmarking

TituLM Bangla Corpus is one of the largest Bangla clean corpus prepared for pretraining, continual pretraining or fine-tuning Large Language Model(LLM) for improving Bangla text generation capability. This dataset contains diverse sources and categories of Bangla text. The largest part of this dataset contains filtered common crawled datasets. As we saw existing all common crawl datasets have issues with proper text extraction from HTML pages and Bangla language specific filtering as all those datasets build for multilingual purposes. Keeping that in mind we applied Trafilatura tool to extract text from common crawl web pages. Compared to existing extraction pages we found this tool perform better. We generate several Bangla language specific quality signals over the dataset and filtered using different quality signals threshold. We also prepared a fine-tuned NLLB model to translate English text to Bangla, and Bangla text to Romanized. We are hoping this dataset contributes to the Bangla research community to buidl more enhance and improved Bangla language model.

Getting Started

To download full datasets:

from datasets import load_dataset

dataset = load_dataset("hishab/titulm-bangla-corpus")

To download a subset:

from datasets import load_dataset

dataset = load_dataset("hishab/titulm-bangla-corpusa", data_dir="<subset_name>")
# example
# dataset = load_dataset("hishab/titulm-bangla-corpusa", data_dir="common_crawl")

Datasets Summary

TituLM Bangla Corpus contains three different categories:

• Common Crawl:
  • Filtered: Contains Common Crawl filtered data. We downloaded the common crawl dump using Athena(Amazon) by Bangla language and language-specific keywords. Then we extract text using Trafilatura which is good tool for web text extraction. We applied several filtering methods. This is the cleanest version of the datasets.
• Translation: Contains Bangla-translated data from English news articles. We used a fine-tuned NLLB model to translate the datasets. In our observation, the fine-tuned nllb model performs better than Google or other available translators. We generated the fine-tuned data using GPT-4 and GPT-4o models.
• Romanized: Contains transliterated Bangla data from Bangla common crawl and news articles. We used a fine-tuned NLLB model to translate the datasets. We generated the fine-tuned data using GPT-4 and GPT-4o models.

Datasets Statistics

• Document counts: Document count describe the total number of document or web pages or page text. For example a news article web page.
• Word Counts: Number of total words counted by basic tokenizer.
• Token Counts: We trained a Tiktoken tokenizer with a large chunk of Bangla text. Here token counts describe the number of tokens counted by https://huggingface.co/hishab/titulm-llama-3.2-3b-v2.0 tokenizer. This tokenizer contains the original Llama 3.1 tokenizer extended with 48k Bangla tokens.

Category	Total Documents (In Millions)	Total Words (In Billions)	Total Tokens (In Billions)
Common Crawl Filtered	24.3	9.94	14.80
Translated	1.74	1.08	1.47
Romanized	5.17	1.89	3.87
Total	31.21	12.91	20.14

Datasets Preparation in Details

Common Crawl

• We used Amazon Athena to query the common crawl datasets. We query by content language, URL host TLD, and dumped the query results.
• We used Trafilatura for extracting text from the query-separated common crawl web HTML pages. We found out that Trafilatura works better to extract text from web HTML pages.
• We generated different quality signals like document word counts, character counts, sentence counts, line ending with terminal punctuations, adult content, etc. We generated a total of 20 quality signals for each document.
• In the final steps, we set a threshold for each quality signal followed by Gopher rule, like word count must be between 50 to 10000, is adult false, sentence count greater than 5, etc. We applied those quality signal thresholds and separated the documents in pass and failed.
• According to our filtering passed percentage 36.76% and failed 62.54%

Translated

• We prepared custom English-to-Bangla translation datasets using OpenAI GPT-4, and GPT-4o models and reviewed the datasets by human annotator.
• We fine-tuned the NLLB model with that dataset and our eval results show promising results on test datasets. Compared to Google Translate our fine-tuned translation seems more natural. We are hoping to publish the model soon.
• Finally, we selected an English newspaper dataset and translated the full dataset to Bangla using the fine-tuned model.

Romanized

• We prepared custom Bangla-to-Romanized Bangla datasets using OpenAI GPT-4, and GPT-4o models and reviewed the datasets by human annotator.
• We fine-tuned the NLLB model with that dataset and our eval results show promising results on test datasets. We are hoping to publish the model soon.
• Finally, we romanized a selected common crawl Bangla dataset using the fine-tuned model.

Citation

@misc{nahin2025titullmsfamilybanglallms,
      title={TituLLMs: A Family of Bangla LLMs with Comprehensive Benchmarking}, 
      author={Shahriar Kabir Nahin and Rabindra Nath Nandi and Sagor Sarker and Quazi Sarwar Muhtaseem and Md Kowsher and Apu Chandraw Shill and Md Ibrahim and Mehadi Hasan Menon and Tareq Al Muntasir and Firoj Alam},
      year={2025},
      eprint={2502.11187},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.11187}, 
}