titulm-llama-3.2-1b-v2.0

Model Information

This model is a continually pretrained version of the meta-llama/Llama-3.2-1B architecture with extended about 42k Bangla tokens, fine-tuned on extensive Bangla datasets. The primary goal of the continual pretraining with token extending was to enhance the model's ability to generate high-quality Bangla text. By extending the pretraining process specifically on Bangla data, the model has demonstrated superior performance in Bangla language understanding evaluation benchmarks and text generation tasks.

Model Architecture: Llama 3.2 is an auto-regressive language model with optimized transformer architecture.

Supported Languages: Bengali (primary) and English (secondary)

Llama 3.2 Model Family: Token counts refer to pretraining data only. All model versions use Grouped-Query Attention (GQA) for improved inference scalability.

Model Release Date: October 24, 2024

Status: This is a static model trained on an offline dataset. Future versions may be released to improve model capabilities.

License: We are using a similar license of Llama 3.2. Use of Llama 3.2 is governed by the Llama 3.2 Community License (a custom, commercial license agreement).

How to use

• Use with transformers

Starting with transformers >= 4.43.0 onward, you can run conversational inference using the Transformers pipeline abstraction or by leveraging the Auto classes with the generate() function.

Make sure to update your transformers installation via pip install --upgrade transformers.

import torch
from transformers import pipeline

model_id = "hishab/titulm-llama-3.2-1b-v2.0"

pipe = pipeline(
    "text-generation", 
    model=model_id, 
    torch_dtype=torch.bfloat16, 
    device_map="auto"
)

pipe("আমাদের দেশের নাম")

Hardware and Software

Training Factors: We used llama-factory training library, Cloud GPU cluster, and production infrastructure for pretraining. Fine-tuning, annotation, and evaluation were also performed on cloud infrastructure.

Training Data

Overview: We have collected a large Bangla raw dataset of text data from a wide variety of sources. Our collected data so far includes a mix of web documents, books, translated text, transliterated text, transcribe text, code-mixed text, conversations, and open sources raw data. The dataset is cleaned and filtered by different filtering criteria to ensure the quality of the data. Our collected data size roughly around 268 GB. Total trained tokens are 37B tokens.

Data sources summary:

• Web documents: Extracted, clean, and filtered common crawl data
• Books: Extracted, clean, filtered books data
• Transcribed text: Used in-house Bangla ASR model to transcribe Bangla audio data
• Translation data: We trained an English-Bangla translation LLM model and used it to translate English data to Bangla
• Code-mixed data: We trained an English-Bangla code-mixed LLM model and used it to generate code-mixed data
• Transliteration data: We trained a Bangla-English transliteration LLM model and used it to generate transliterated data
• Synthetic data: We generated synthetic data using a Bangla LLM model
• Others: We scrapped some selected website data, used open-source data, and used some other data sources

Token Extending

We trained a separate Bangla tokenizer using Tiktoken library on 48 GB Bangla datasets (sampled from main pretraining data) with a vocabulary size 48k and separated 42k tokens for adding with the pretrained model. We extended the model's vocabulary with these tokens and continued the pretraining process on Bangla data. The token extending process was done to enhance the model's ability to generate high-quality Bangla text. Our updated vocab size is 170k whereas the original llama-3.2 vocab size is 128k.

Benchmarks - Bangla Text

In this section, we report the results for titulm-llama-3.2-1b-v2.0 models on standard automatic benchmarks. For all these evaluations, we used lm-evaluation-harness evaluations library.

Evaluation Datasets

We evaluated our pretrained models on both Bangla and English benchmark datasets. Although the model is trained on Bangla data, its English capability is also evaluated on English benchmark datasets. The evaluation datasets are as follows:

Bangla Benchmark datasets

We evaluated the models on the following datasets:

• Bangla MMLU: A private multiple choice question dataset developed by Hishab curated from various sources.
• CommonsenseQa Bangla: A Bangla translation of the CommonsenseQA dataset. The dataset was translated using a new method called Expressive Semantic Translation (EST), which combines Google Machine Translation with LLM-based rewriting modifications.
• OpenbookQA Bangla: A Bangla translation of the OpenbookQA dataset. The dataset was translated using a new method called Expressive Semantic Translation (EST), which combines Google Machine Translation with LLM-based rewriting modifications.
• Piqa Bangla: A Bangla translation of the Piqa dataset. The dataset was translated using a new method called Expressive Semantic Translation (EST), which combines Google Machine Translation with LLM-based rewriting modifications.
• BoolQ Bangla: The dataset contains 15,942 examples, with each entry consisting of a triplet: (question, passage, answer). The questions are naturally occurring, generated from unprompted and unconstrained settings. Input passages were sourced from Bangla Wikipedia, Banglapedia, and News Articles, and GPT-4 was used to generate corresponding yes/no questions with answers.

English Benchmark datasets

• MMLU: This is a massive multitask test consisting of multiple-choice questions from various branches of knowledge.
• CommonseQa: CommonsenseQA is a new multiple-choice question-answering dataset that requires different types of commonsense knowledge to predict the correct answers.
• OpenbookQA: OpenBookQA aims to promote research in advanced question-answering, probing a deeper understanding of both the topic (with salient facts summarized as an open book, also provided with the dataset) and the language it is expressed in.
• Piqa: The PIQA dataset focuses on physical commonsense reasoning, challenging AI to handle everyday situations requiring practical knowledge and unconventional solutions. Inspired by instructables.com, it aims to enhance AI's ability to understand and reason about physical interactions.
• BoolQ: BoolQ is a question answering dataset for yes/no questions containing 15942 examples. These questions are naturally occurring ---they are generated in unprompted and unconstrained settings. Each example is a triplet of (question, passage, answer), with the title of the page as optional additional context. The text-pair classification setup is similar to existing natural language inference tasks.

Evaluation Results

Evaluation on Bangla Benchmark datasets

• However, titulm-llama-3.2-1b-v2.0 performs better in Commonsense QA and PIQA BN, where it outperforms the original model in both 0-shot and 5-shot settings.
• llama-3.2-1b generally outperforms titulm-llama-3.2-1b-v2.0 on Bangla datasets, especially in the Bangla MMLU and BoolQ BN tasks, achieving a higher score in the 0-shot setting.
• The models perform similarly on OpenBook QA, with marginal differences.

Evaluation on English Benchmark datasets

• llama-3.2-1b shows consistently better performance on English datasets, especially in tasks like MMLU, BoolQ, Commonsense QA, and PIQA.
• In comparison, titulm-llama-3.2-1b-v2.0 underperforms in both 0-shot and 5-shot settings across all tasks in English.
• It was expected as we trained the model only on Bangla datasets.

Instruction Tuned Models

Intended Use

• Bangla text generation
• Bangla language understanding tasks
• Bangla instruction fine-tuning tasks