README.md

metadata

language:
  - multilingual
  - en
  - sw
  - ha
  - yo
  - ig
  - zu
  - sn
  - ar
  - am
  - fr
  - pt
tags:
  - zero-shot-image-classification
  - image generation
  - visual qa
  - text-image embedding
  - image-text embedding
  - pytorch
  - sartify
  - visual conversional ai
  - image semantic retrival
  - african raw resourced languages
  - safetensors
  - vision-text-dual-encoder
license: apache-2.0
library_name: transformers
datasets:
  - sartifyllc/AViLaData
metrics:
  - accuracy
  - code_eval
  - bleu

AViLaMa : African Vision-Languages Aligment Pre-Training Model.

Learning Visual Concepts Directly From African Languages Supervision. Paper is coming

Model Details

AViLaMa is the large open-source text-vision alignment pre-training model in African languages. It brings a way to learn visual concepts directly from African languages supervision. Inspired from OpenAI CLIP, but with more based on African languages to capture the nuances, cultural context, and social aspect use of our languages that are so impossible to get just from machine translation. It includes techniques like agnostic languages encoding, data filtering network etc... All for more than 12 African languages, trained on the #AViLaDa-2B datasets of filtered image-text pairs.

• Developed by : Sartify LLC (www.sartify.com)
• Authors : Sartify LLC Research Team
• Funded by : Sartify LLC, Open Source Community, etc..(We always welcome other donors)
• Model type : multilingual & multimodality transformer
• Language(s) : en (English), sw (Swahili), ha (Hausa), yo (Yoruba), ig (Igbo), zu (Zulu), sn (Shona), ar (Arabic), am (Amharic), fr (French), pt (Portuguese)
• License: apache 2.0

Load model from hugging face.

import torch
from transformers import AutoModel, AutoTokenizer

model = AutoModel.from_pretrained("sartifyllc/AViLaMa")
tokenizer = AutoTokenizer.from_pretrained("sartifyllc/AViLaMa")
model = model.eval()

Model Sources

• Repository : AViLaMa-Sources
• Datasets : Coming...
• Paper : Coming...
• Demo : Coming...

Direct & Downstream Use In African Languages:

1. zero shot semantic image retrieval and ranking tasks.
2. zero shot image classification tasks.
3. visual QA tasks with African languages.
4. visual conversional GenAI tasks.
5. image capturing tasks.
6. images and art generation guiding and conditioning tasks.
7. text-images analysis tasks.
8. content moderation task etc....

Citation

BibTeX:

AViLaMa paper
@article{sartifyllc2023africanvision,
  title={AViLaMa: Learning Visual Concepts Directly From African Languages Supervision},
  author={Sartify LLC Research Team},  
  journal={To be inserted},
  year={2024}
}