lookarooka/looka-Stock-Base-8B

lookarooka/looka-Stock-Base-8B: Specialized Financial Intelligence

This model, developed by lookarooka, is an 8 billion parameter language model built upon the unsloth/llama-3-8b base, enhanced with a Korean financial and stock-specific knowledge model (Bllossom family). The primary goal was to create a highly specialized model for financial and stock-related inquiries, minimizing general conversation and hallucinations.

Development and Optimization

Initial experiments involved a "task_sub" operation to mathematically subtract the base llama-3-8b model's general knowledge, aiming to sharpen its financial expertise. However, these attempts, even at low intensity, led to critical issues such as corrupted Korean sentence structures, infinite loops, and foreign language outputs. It was determined that this subtraction method damaged essential tokens responsible for controlling Korean grammar.

Final Approach and Key Features

Consequently, the developers opted to use the original combined model (My-Stock-Base-8B) without further subtraction, preserving its full knowledge and linguistic integrity. To ensure efficient performance on personal computers (e.g., LM Studio, Connect AI), the model has been optimized through 4-bit quantization (Q4_K_M GGUF). Users are advised to control hallucinations and general chatter effectively via system prompt settings rather than through model modification.

Good for:

• Specialized financial and stock market analysis
• Answering specific questions related to Korean finance
• Deployment on personal computers with optimized performance