Spatial-LLaVA-7B Model Card
🤖 Model details
Model type:
This finetuned LLaVA model is trained from
liuhaotian/llava-pretrain-vicuna-7b-v1.3 for improving spatial relation reasoning of large multi-modal model.
LLaVA is an open-source chatbot trained by fine-tuning LLaMA/Vicuna on GPT-generated multimodal instruction-following data.
It is an auto-regressive language model, based on the transformer architecture.
🎯 Intended use
Primary intended uses:
The primary use of LLaVA 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
📊 Evaluation
A collection of 10 benchmarks:
| Model | VQAv2 | GQA | VizWiz | SQA | TextVQA | POPE | MME | MM-Bench | MM-Bench-cn | MM-Vet |
|---|
| LLaVA-1.5-7b | 78.5 | 62.0 | 50.0 | 66.8 | 58.2 | 85.9 | 1510.7 | 64.3 | 58.3 | 31.1 |
| Spatial-LLaVA-7b | 79.7 | 62.7 | 48.7 | 68.7 | 58.5 | 87.2 | 1472.7 | 67.8 | 60.7 | 31.6 |
Qualitative Spatial Relations
| Model | Below/Above | Left/Right | Big/Small | Tall/Short | Wide/Thin | Behind/Front | Avg |
|---|
| LLaVA-1.5-7b | 53.91 | 53.49 | 45.36 | 40.00 | 50.00 | 51.04 | 48.97 |
| LLaVA-1.5-13b | 54.28 | 52.32 | 45.36 | 48.57 | 49.02 | 47.92 | 49.67 |
| Spatial-LLaVA-7b | 56.32 | 66.28 | 60.82 | 48.57 | 49.02 | 52.08 | 55.12 |
Quantitative Spatial Relations
| Model | Direct Dist (m / ratio) | Horizontal Dist (m / ratio) | Vertical Dist (m / ratio) | Width (m / ratio) | Height (m / ratio) | Direction (° / ratio) |
|---|
| LLaVA-1.5-7b | 12.90 / 1.06 | 10.68 / 2.03 | 20.79 / 0.94 | 24.19 / 0.50 | 14.29 / 5.27 | 10.23 / 58.33 |
| LLaVA-1.5-13b | 13.71 / 0.93 | 10.68 / 3.56 | 16.83 / 0.85 | 15.32 / 0.57 | 17.67 / 5.8 | 14.77 / 54.29 |
| Spatial-LLaVA-7b | 24.19 / 0.57 | 14.56 / 0.62 | 41.58 / 0.42 | 22.58 / 1.12 | 18.25 / 2.92 | 20.45 / 56.47 |
🙏 Acknowledgements
We thank Liu Haotian et al. for the LLaVA pretrained script, weights and LLaVA-v1.5 mixture dataset; the teams behind CLEVR, TextCaps, VisualMRC and VQAv2 (via “HuggingFaceM4/the_cauldron”); remyxai for OpenSpaces; Anjie Cheng et al. for Spatial-Bench and data pipeline; Google for OpenImages; and Hugging Face for their datasets infrastructure.
