|
--- |
|
license: apache-2.0 |
|
datasets: |
|
- tabtoyou/KoLLaVA-Instruct-150k |
|
- tabtoyou/KoLLaVA-CC3M-Pretrain-595K |
|
language: |
|
- ko |
|
library_name: transformers |
|
tags: |
|
- LLaVA |
|
- KoVicuna |
|
- KoLLaVA |
|
- KoAlpaca |
|
- CLIP |
|
--- |
|
|
|
# KoLLaVA : Korean Large Language and Vision Assistant (feat. LLaVA) |
|
This model is a large multimodal model (LMM) that combines the LLM([KoVicuna](https://huggingface.co/junelee/ko_vicuna_7b)) with visual encoder of CLIP([ViT-14](https://huggingface.co/openai/clip-vit-large-patch14)), trained on [Korean visual-instruction dataset](https://huggingface.co/datasets/tabtoyou/KoLLaVA-Instruct-150k). |
|
|
|
Detail codes are available at [KoLLaVA github repository](https://github.com/tabtoyou/KoLLaVA) |
|
|
|
### Training hyperparameters |
|
* learning rate : 2e-5 |
|
* train_batch_size: 16 |
|
* distributed_type: multi-GPU (A100 80G) |
|
* num_devices: 4 |
|
* gradient_accumulation_steps: 1 |
|
* total_train_batch_size: 64 |
|
* total_eval_batch_size: 16 |
|
* lr_scheduler_type: cosine |
|
* num_epochs: 1 |
|
|
|
Model License: Apache License 2.0 |