LLAMA-3-Korean-Bllossom-8B

Introduction

The LLAMA-3-Korean-Bllossom-8B is a Korean-English bilingual language model based on the open-source LLama3. It is designed to enhance the connection between Korean and English knowledge and is fine-tuned with custom data to suit Korean language and culture. It supports advanced text generation, conversational tasks, and inference endpoints.

Architecture

This model is built on the Meta-Llama-3-8B architecture and supports both English and Korean languages. It utilizes the transformers library and is optimized for efficient performance, including a vocabulary expansion for Korean expressiveness and extended context length.

Training

The model underwent substantial pre-training, with a dataset size increased to 250GB. It incorporates knowledge linking between Korean and English and uses human feedback for further refinement. Additionally, the model aligns vision and language components, utilizing vision transformer integration.

Guide: Running Locally

1. Install Dependencies: Use the command pip install torch transformers==4.40.0 accelerate to set up the required libraries.
2. Load the Model: Use the transformers library to load the model and tokenizer. Example code snippets are provided to facilitate this.
3. Inference: Execute text generation using either the pipeline or AutoModel approach.
4. Environment: While local execution is possible, using cloud GPUs like Google Colab is recommended to leverage free GPU resources or to deploy a quantized model on CPU. A Colab tutorial is available for guidance.

License

The model is released under the LLama3 license, permitting both academic and commercial use.