Create README.md

README.md

@@ -0,0 +1,44 @@
+---
+license: apache-2.0
+datasets:
+- aldente0630/LLaVA-Pretrain-Ko
+- tabtoyou/KoLLaVA-v1.5-Instruct-581k
+language:
+- ko
+- en
+base_model:
+- jjhsnail0822/danube-ko-1.8b-base
+- openai/clip-vit-large-patch14-336
+tags:
+- h2o-danube2
+- korean
+- sLLM
+- llm
+- multimodal
+- LLaVA
+---
+
+
+## Model Details
+
+llava-danube-ko-1.8b-instruct is a Korean multi-modal language model based on [jjhsnail0822/danube-ko-1.8b-base](https://huggingface.co/jjhsnail0822/danube-ko-1.8b-base).
+
+## Model Developers
+
+Jinhong Jeong, Ungsang Yoon
+
+## Model Architecture
+
+We used the [LLaVA-NeXT](https://github.com/LLaVA-VL/LLaVA-NeXT) technique to add multi-modal capabilities to the model by 2-stage visual instruction tuning. [jjhsnail0822/danube-ko-1.8b-base](https://huggingface.co/jjhsnail0822/danube-ko-1.8b-base) was used as the base LLM, and the visual encoder was fine-tuned from [openai/clip-vit-large-patch14-336](https://huggingface.co/openai/clip-vit-large-patch14-336). The model has sequence length of 2048. 
+
+## Training Datasets
+
+In the multi-modal pretrain stage, images filtered from LAION/CC/SBU dataset was used. For the visual instruction tuning stage, we prepared the training dataset from COCO, GQA, Visual Genome datasets and EKVQA dataset from AI-Hub. About 90GB of compressed image data was used for the whole training process.
+
+## Model Benchmark
+
+TBA
+
+## Disclaimer
+
+The Model can generate information that is biased, discriminatory, socially inappropriate, etc. The Model can also generate information that is not accurate. The Model is used at your own risk, and the developers are not responsible for the information generated by the model.