Saxo/Linkbricks-Horizon-AI-Korean-llama3-sft-dpo-8b-base

Model Overview

Saxo/Linkbricks-Horizon-AI-Korean-llama3-sft-dpo-8b-base is an 8 billion parameter language model developed by Dr. Yunsung Ji (Saxo), a data scientist at Linkbricks, an AI and big data analytics company. It is built upon the meta-llama/Meta-Llama-3-8B base model and has undergone specialized Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO) for Korean language processing.

Key Training Details

• Base Model: Meta-Llama-3-8B
• Training Hardware: 8 H100-60G GPUs on Google Cloud Platform (GCP)
• Training Duration: 4 hours of instructional training
• Training Data: 8000 tokens, with a focus on Korean content.
• Libraries Used: Accelerate, Deepspeed Zero-3 for efficient training.
• Tokenizer: Utilizes the original Llama 3 tokenizer; no Korean vocabulary expansion has been applied in this version. Linkbricks offers separate models with over 200,000 Korean-specific vocabulary tokens upon request.

Configuration Highlights

The model was trained with BitsAndBytesConfig for 4-bit quantization, using nf4 quantization type and paged_adamw_32bit optimizer. Training involved 20 epochs with a learning rate of 2e-4 and gradient accumulation steps of 4, supporting bf16 where available.

Intended Use

This model is designed for applications requiring a Llama 3-based LLM with enhanced capabilities for Korean language understanding and generation, particularly in scenarios where SFT and DPO fine-tuning on Korean data are beneficial.