Skip to content

hishab-nlp/titulm

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

4 Commits
 
 
 
 
 
 
 
 

Repository files navigation

TituLM

Hugging Face

TituLM is a collection of open source LLMs trained by Hishab for the purpose of better language understanding and generation capabilities. At Hishab, we are pushing the boundaries of what is possible with LLMs for product development. According to this pipeline we are pretraining and finetuning models on a variety of tasks to improve the capabilities of the models. Although TituLM is not bound to any specific language but it mostly focuses on Bangla language.

Models

TituLM Llama Family

We have trained multiple variants of Llama 3.2 family models with different sizes and configurations. Our released models are:

3B Family

1B Family

TituLM Gemma Family

We have trained multiple variants of Gemma family models with different sizes and configurations. Our released models are:

Gemma 2 2B

TituLM MPT Family

  • titulm-mpt-1b-v1.0: Trained with 4.51B Bangla tokens with custom Bangla tokenizer.
  • titulm-mpt-1b-v2.0: Trained with 43B Bangla and English tokens with custom Bangla+English tokenizer.

Usage

Generation using transformers

# pip install transformers
import torch
from transformers import pipeline

model_id = "hishab/titulm-llama-3.2-3b-v2.0"

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

pipe("আমাদের দেশের নাম")

Benchmark

  • Clone the forked version of lm-evaluation-harness repository.
  • Now run the following commands to evaluate the models.
  • Pass required arguments according to your needs.
git clone https://github.com/hishab-nlp/lm-evaluation-harness.git
cd lm-evaluation-harness
pip install -e .


# Running the benchmark
# https://github.com/hishab-nlp/lm-evaluation-harness/blob/main/scripts/bangla_lm_benchmark.py
cd scripts
# pass arguemnt according to your needs
python bangla_lm_benchmark.py

About

TituLM Development Library

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages