qlora-llama-3-8b

AIPI 590 Large Language Models

Project 1 - Fine Tuning LLM

Files:

• model.ipynb
  • notebook containing the code for fine tuning the Llama 3 model using QLoRa
• data/train.json
  • json file containing the training set provided in the FINQA paper
• data/test.json
  • json file containing the validation set provided in the FINQA paper

Process:

The focal property of interest is analysis financial documents for numerical reasoning. Specifically numerical reasoning over quarterly financial filings with the SEC. The Llama-3-8B model was chosen to fine tune using the QLoRa approach. This approach was chosen due to the paper's findings of a performance increase while utilizing minimal memory and hardware. The aggressive quantization seemed to significantly decreased training time while offering increased performance on financial analysis.

Evaluation:

Rouge Score

Collaborators:

• Keese Phillips

Attribution:

• FINQA: A Dataset of Numerical Reasoning over Financial Data
• LORA: Low-Rank Adaptation of Large Language Models
• QLORA: Efficient Finetuning of Quantized LLMs