metadata
license: gemma
language:
- en
pipeline_tag: image-text-to-text
Cerule - A Tiny Mighty Vision Model
Based on Google's - Gemma-2b + SigLIP
ββββββββββββββββββββββ βββ ββββββ ββββββββ
βββββββββββββββββββββββββββ ββββββ ββββββββ
βββ ββββββ βββββββββββ ββββββ ββββββ
βββ ββββββ βββββββββββ ββββββ ββββββ
βββββββββββββββββββ ββββββββββββββββββββββββββββ
ββββββββββββββββββ βββ βββββββ ββββββββββββββββ
We train and release "Cerule", a tiny yet powerful Vision Lanuage Model based on the newly released Google's Gemma-2b and Google's SigLIP.
- Pretraining stage : 650K images (A LAION Subset)
- Finetuning stage : 695K images (SVIT-mix-665K - Bunny mix modified by BAAI)
The training setup was 4xA100's 80GB
and took ~6 hours to pretrain and ~13 hours to finetune. We modify and adapt the training code from Bunny.
Installing requirements
pip install -qr https://huggingface.co/Tensoic/Cerule-v0.1/resolve/main/requirements.txt
Training:
Training code Released !!! https://github.com/tensoic/Cerule
Inference:
Clone the following repo and following instructions for a CLI based inference. https://github.com/tensoic/Cerule
License
Model subject to Gemma(base model license) terms of use along with the underlying datasets(LAOIN and SVIT) subject to their respective licenses. All codes are Apache 2.0
Acknowledgments
We sincerely thank the Amazing teams at Google, LLaVA, and BAAI without which this project would not have been possible!