Financial Content Social Media Generator

An end-to-end machine learning pipeline designed to automate the creation of professional social media posts (LinkedIn, X, etc.) from financial articles. This project explores different LLM adaptation techniques to match a specific professional "voice" while maintaining factual accuracy.

Project Overview

The pipeline transforms a raw URL into a series of social media post variations through four main stages:

1. Automated Scraping: Uses Playwright to extract core article text while filtering out legal boilerplate, disclosures, and navigation links.
2. Summarization: Implements a BART-large-cnn model to condense long-form financial articles into concise summaries.
3. Model Adaptation: Compares three distinct approaches to generate high-quality posts:
   • Zero-Shot/Few-Shot Prompting: Direct instructions to the base model (Llama-3.2-1B-Instruct).
   • Prompt-Tuning: Training virtual tokens to guide style.
   • LoRA (Low-Rank Adaptation): Efficiently fine-tuning model weights for specific tonal alignment.
4. Evaluation: Uses ROUGE-L metrics to quantitatively measure semantic overlap between AI-generated posts and professional human-written examples.

Key Findings

• LoRA Performance: LoRA fine-tuning was the most effective method, producing the most concise and stylistically accurate posts.
• Metric Analysis: While BLEU was found to be too strict for creative rephrasing, ROUGE-L proved to be a reliable metric for capturing semantic similarity.
• Quantitative Results:
  • Base Model ROUGE-L: ~0.092
  • Prompt-Tuned ROUGE-L: ~0.093
  • LoRA Fine-Tuned ROUGE-L: ~0.107

Repository Structure

• data_processing.py: Handles web scraping, text cleaning, and BART-based summarization.
• lora_tuning.py: Script for fine-tuning Llama-3.2-1B-Instruct using LoRA.
• prompt_tuning.py: Script for training virtual tokens (Prompt-Tuning).
• post_generation.py: Generates social media variations and runs ROUGE evaluation.

Usage

1. Preprocess Data

Scrape content and generate summaries from a list of URLs:

python data_processing.py input_data.xlsx

2. Training the Models

You can train the model using two different Parameter-Efficient Fine-Tuning (PEFT) methods:

• LoRA Fine-Tuning: Run the tuning script using the preprocessed data to fine-tune the base model with LoRA.

  python lora_tuning.py data_path output_path epochs lr batch_size lora_r lora_alpha lora_dropout

• Prompt Tuning: Run the prompt tuning script using the preprocessed data to train virtual tokens on the model.

  python prompt_tuning.py data_path output_path num_vir_tokens num_epochs lr batch_size

3. Generating Content

Run the generation script to create social media posts from summaries:

python post_generation.py model_path data_path output_path num_posts temp top_p max_new_tokens

Use --analyze if you wish to evaluate generated posts using the ROUGE metric.

4. Customization and Iteration

The scripts are designed to be easy to iterate with.

• You can run any script with -h to see additional parameters for customizing the models and generation process.
• Adjust the number of posts, temperature, top-p sampling, and maximum tokens using the command-line arguments to explore different outputs.