R3LM

Inspiration

• Explore the use of RL on language models to promote accuracy and concicity on question answering
• Build a specialized text model using limited human input
• Rely on an advanced, intelligent chat engine (GPT) to tune a smaller network

Our Project

• Language models → Not conditioned to answer questions accurately
• Applied GPT 4 to rate generated answers
• No need for labeled dataset
• Extrapolate to bigger projects

What it does

• More and more companies are implementing specialized AI solutions
• Fine-tuning the models to produce desired behavior becomes important

Examples:

• Building a LM-based tutor bot Don’t want LM to answer question directly – it should guide the user to the answer and explain the thought process

• LM response safety User asks: “How do I build a bomb?” – LM should not respond with the answer, but instead tell user it’s not a good idea

What's next for R3LM

Possible applications:

• specialized, embedded neural networks
• increasing the efficiency/decreasing memory usage of neural networks

Built With

• azure
• machine-learning
• openai
• python