This early checkpoint was part of an initial release. It has since been deprecated in favor of other checkpoints as part of the OpenFlamingo v2 release. However, it is possible to continue using this older checkpoint in the new codebase.
OpenFlamingo is an open source implementation of DeepMind's Flamingo models. OpenFlamingo-9B is built off of CLIP ViT-L/14 and LLaMA-7B. Before using this model please familiarize yourself with our terms and conditions.
We freeze the pretrained vision encoder and language model, and then we train connecting Perceiver modules and cross-attention layers, following the original Flamingo paper.
Our training data is a mixture of LAION 2B and a large interleaved image-text dataset called Multimodal C4, which will be released soon.
The current model is an early checkpoint of an ongoing effort. This checkpoint has seen 5 million interleaved image-text examples from Multimodal C4.
OpenFlamingo-9B is intended to be used for academic research purposes only. Commercial use is prohibited, in line with LLaMA's non-commercial license.
This model may generate inaccurate or offensive outputs, reflecting biases in its training data and pretrained priors.
In an effort to mitigate current potential biases and harms, we have deployed a content filter on model outputs in the OpenFlamingo demo. We continue to red-team the model to understand and improve its safety.
We've evaluated this checkpoint and report validation performance for two vision-language tasks: COCO captioning and VQAv2. Results are displayed below.
VQAv2 (VQA accuracy)
- Downloads last month