SharkTankAI

This project aims to fine-tune a language model on Shark Tank pitch transcripts to generate new, creative business pitches.

Project Structure

• data-generation/: Contains scripts for collecting YouTube video URLs, transcribing them, and preparing the data for fine-tuning.
• fine-tuning/: Contains scripts for fine-tuning a language model (TinyLlama) using PEFT (LoRA) and merging the LoRA adapters with the base model.
• transcripts/: Stores the transcribed text from YouTube videos.
• lora-llama/: Contains the checkpoints from the fine-tuning process.
• merged_model/: Stores the final merged language model.

Setup

1. Clone the repository:

   git clone https://github.com/your-username/SharkTankAI.git
   cd SharkTankAI

2. Create a virtual environment and install dependencies:

   python -m venv .venv
   source .venv/bin/activate
   pip install -r requirements.txt

Usage

1. Data Collection and Preparation

First, collect video URLs, transcribe them, and generate the JSON dataset:

python data-generation/collect_videos.py
python data-generation/transcribe_videos.py
python data-generation/generate_json.py

This will:

• Search YouTube for 'Shark Tank' videos and save URLs to video_urls.txt.
• Transcribe the videos and save transcripts to the transcripts/ directory.
• Combine all transcripts into shark_tank_pitches.json.

2. Fine-tuning the Model

Once the data is prepared, you can fine-tune the TinyLlama model:

python fine-tuning/fine_tune.py

This script will save the fine-tuned LoRA adapters in the lora-llama/ directory.

3. Merging the LoRA Adapters

After fine-tuning, merge the LoRA adapters with the base model to create a standalone model:

python fine-tuning/merge_lora.py

The merged model will be saved in the merged_model/ directory.

Model Details

The project uses TinyLlama/TinyLlama-1.1B-Chat-v1.0 as the base model for fine-tuning. The model has been trained on Shark Tank transcripts to enable it to act as a 'negator model', providing critical analysis of business pitches. The fine-tuned model is then converted to GGUF format and can be merged with Ollama for local inference.

Future Improvements

• Implement a script to generate new pitches using the merged model.
• Explore different base models and fine-tuning parameters.
• Add more robust error handling and logging.