Model Overview

The kmseong/SN-GSM8K-SFT-Model is a specialized Llama 3.1 8B Instruct model, fine-tuned by kmseong to excel in mathematical reasoning. It leverages Low-Rank Adaptation (LoRA) with a rank of 8 and alpha of 16, targeting key attention and feed-forward modules (q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj).

Key Capabilities

• Enhanced Mathematical Reasoning: Specifically trained on the GSM8K dataset to improve performance on grade school math word problems.
• Llama 3.1 8B Instruct Base: Built upon the robust meta-llama/Llama-3.1-8B-Instruct foundation.
• LoRA Fine-tuning: Efficiently fine-tuned with 500 training samples over 3 epochs, achieving a final training loss of approximately 0.43.

Performance and Use Cases

This model demonstrates a 55.00% accuracy on the GSM8K test set (11 out of 20 samples), indicating its focused improvement in mathematical problem-solving. It is particularly well-suited for:

• Solving Math Word Problems: Designed to generate step-by-step reasoning and provide numerical answers for arithmetic and algebraic problems.
• Educational Tools: Can be integrated into applications that assist with learning or practicing mathematics.
• Research in Mathematical Reasoning: Provides a fine-tuned base for further experimentation or analysis of LLM capabilities in mathematics.

Users should note that while optimized for math, its performance on other general tasks may vary. The model can be easily integrated using Hugging Face Transformers and PEFT libraries, with options to load the LoRA adapter or merge its weights into the base model for inference.