|
--- |
|
license: apache-2.0 |
|
--- |
|
|
|
# ASMv2 Model Card |
|
|
|
## Model details |
|
|
|
**Model type:** |
|
ASMv2 is an open-source chatbot trained by fine-tuning LLaMA/Vicuna on multimodal instruction-following data. |
|
It integrates the Relation Conversation (REC) ability while maintaining powerful general capabilities. |
|
This model is also endowed with grounding and referring capabilities, exhibiting state-of-the-art performance on region-level tasks, and can be naturally adapted to the Scene Graph Generation task in an open-ended manner. |
|
|
|
**Model date:** |
|
ASMv2 was trained in January 2024. |
|
|
|
**Paper or resources for more information:** |
|
https://github.com/OpenGVLab/all-seeing |
|
|
|
## License |
|
ASMv2 is open-sourced under the Apache License 2.0. |
|
|
|
**Where to send questions or comments about the model:** |
|
https://github.com/OpenGVLab/all-seeing/issues |
|
|
|
## Intended use |
|
**Primary intended uses:** |
|
The primary use of ASMv2 is research on large multimodal models and chatbots. |
|
|
|
**Primary intended users:** |
|
The primary intended users of the model are researchers and hobbyists in computer vision, natural language processing, machine learning, and artificial intelligence. |
|
|
|
## Training dataset |
|
The pretrain phase employs [5M filtered samples](https://storage.googleapis.com/sfr-vision-language-research/BLIP/datasets/ccs_filtered.json) from CC12M, [10M filtered samples](https://huggingface.co/datasets/Weiyun1025/AS-V2/blob/main/as_pretrain_10m.json) from AS-1B, and 15M filtered samples from [GRiT](https://huggingface.co/datasets/zzliang/GRIT). |
|
|
|
The instruction-tuning phase employs [4M samples](https://huggingface.co/datasets/Weiyun1025/AS-V2/blob/main/as_mix_4m.json) collected from a variety of sources, including image-level datasets |
|
|
|
See [here](https://github.com/OpenGVLab/all-seeing/tree/main/all-seeing-v2#training) for more details. |
|
|
|
## Evaluation dataset |
|
A collection of 20 benchmarks, including 5 academic VQA benchmarks, 7 multimodal benchmarks specifically proposed for instruction-following LMMs, 3 referring expression comprehension benchmarks, 2 region captioning benchmarks, 1 referring question answering benchmark, 1 scene graph generation benchmark, and 1 relation comprehension benchmark. |
