Team Name: Sigolt
DunXr - Innovative storytelling with the help of GenAI
A fun way to play a world-building storytelling game where AI helps you generate a storyline and is your partner in navigating through a complex landscape. With choices to make at every turn, will your decision end up being the right one?
We used the Llama-2-7b model for this purpose. Training was conducted using a custom-made dataset with the help of sources that included datasets of fictional stories, along with some custom storylines that were generated by generative AI. After this, the model was able to create a decent storyline with the given prompt, followed by two choices that would influence the future.
We used the foundational model to be Llama-2 with 7 billion parameters. The implemetation of Intel's optimization of PyTorch (IPEX) greatly decreased the time taken for training the model. The guidance provided for fine-tuning the model was especially helpful and was greatly appreciated as our team was very new to LLMs and AI in general.
The Intel Developer Cloud offered a fantastic platform with access to high-speed computing power which helped us experiment and learn a lot as we built our first LLM-based project. This platform showed us that faster training was achievable during our experimentation.
Initially, the game begins with either the player setting the world type and game type (Horror, Adventure and so on) or the AI generating a randoms storyline with the theme set by the player. After that, based on a dice roll, the one with the greater dice roll decides each part of the story. When a natural end occurs, if the player's dice roll wins, the game is over and the player wins. Else, the player loses. All throughout the game, the next choices are generated by the AI, with each choice leading to completely different scenarios.
- NLP: Learnt more about text generation and prompt generation.
- Intel Technology: Learnt more about Intel's OneAPI toolkit and Intel Developer Cloud. Experimented with IPEX for highly optimized fine-tuning.
- Frontend: NextJS, ReactJS
- Backend: IDC, Jupyter Notebook, Intel AI Analytic Toolkit, FastAPI
- LLM: Llama-2-7b