Inspiration
Our inspiration came from noting that getting detailed and relevant output from LLMs can be quite a trying matter. The LLM will often give irrelevant output when fed too much information, but likewise won't give specific enough output when fed too little information.
What it does
The solution we built takes long-form prompts written in natural language and mutates them in an attempt to create better prompts. This blind prompt engineering creates a population of these mutated prompts and uses a fitness function to select the best.
How we built it
This program was built using Python.
Challenges we ran into
The key challenges we ran into were:
- extracting sentences correctly for the mutation function.
- Finding the good way to select sentences for mutation.
- crossing over prompts within the population.
- evaluating prompts.
- Large computational requirements of the algorithm due to LLMs as well as longer prompts.
Accomplishments that we're proud of
The entire journey has been amazing! We didn't had any knowledge of LLMs and NLP, but learned throughout the event. There has been a lot of great milestones that we are proud of but the key ones would be,
- Learning about LLMs and emerging field of Prompt Engineering
- Learning to use pretrained models for generating responses
- Learning about Genetic Algorithms
- Finishing this journey with everyone!
What's next for Linguistic Genetic Bandits
Towards a new beginning
We are going to develop the solution further to make it more usable by the open source community. Further improving the algorithm through,
- Exploring mutation of a sentence based on previous and next sentences to capture more structure in the prompt during the rephrasing.
- Fine tuning the parameters for some specific tasks, to provide pre built prompts for the tasks.
- Continue testing and improving efficiency with the hopes that it might be able to benefit people within the open source community in the future.
- Explore more sophisticated Reinforcement learning algorithms to further improve the selection of sentences for mutation.
- Explore more strategies for crossover in NLP to optimize the crossover operation in the algorithm.
- Explore other algorithms belonging to the evolutionary algorithms to translate the key improvements to specific problems
Log in or sign up for Devpost to join the conversation.