My AI powered D&D random encounter generator is now online – and it’s free to use! Check out Random Encounters AI or read below about how I created the website. I’ve spent a lot of time perfecting the process – and it’s an ongoing effort – but I’m happy with the results so far and have received quite a bit of love for the website, especially making it free to use. I hope you enjoy reading about how I created by D&D random encounter generators with GPT-3 from OpenAI. I’ve been working with D&D Random Encounters for a while, and artificial intelligence and natural language processing (NLP) has helped tremendously.
Studying Random Encounter Generators
Before I tackled using GPT-3 from OpenAI to create random encounters, I did a bit of research. A lot of great procedural generation tools exist online, but not many were using language models like GPT-3 for tabletop RPG content like random encounter tables. I saw this as an opportunity to experiment with GPT-3 in a table top RPG context.
With my experience using GPT-3 for character backstory generation (and a lot more), I collected data about how I would want to set up a random encounter generator for D&D and other tabletop roleplaying games. By the time I finished my research, I had a massive list of potential features and data I would want to use to generate random encounters.
Compiling Random Encounter Examples
Rather than using random encounters from other authors, I began to use the GPT-3 Playground to generate my own random encounter examples. For months, I generated thousands of random encounters for a variety of locations, including mountains, swamps, forests, cities, and much more. I paid a lot of attention to editing the output of the AI.
During this phase, I wrote a lot of my own examples in order to allow GPT-3 to emulate my writing style. I did this because I wanted the AI to produce random encounters with a similar style to what I would write. This made it easier to spot errors in my data and to edit the output so that it more closely matched my expectations. Also, it meant I wasn’t stealing work from other authors.
Fine-Tuning GPT-3 Davinci Language Model
Last month, I started fine-tuning GPT-3 Curie model with 6,000+ individual random encounters. The first few fine-tunes were a bit rocky, but I gradually pruned and added to the dataset of random encounter examples. After doing this, I contacted OpenAI and asked (begged) for access to fine-tuning their best model – Davinci.
(Note: At this time, fine-tuning is a lot easier at OpenAI, and it’s also easier to get an account.)
Because I wanted to output 10 random encounters at a time, I trained with 600+ sets of 10 random encounters bundled together for a particular location. I had much better results when fine-tuning Davinci with 4,200+ character backstory examples, but even 600+ random encounter examples was enough to start getting better output for a cheaper price. Good news!
Launching the Random Encounters AI Website
Because my random encounter generator is still not perfect and it costs between $0.05 and $0.10 per batch of 10 random encounters, I decided to create a website where I could highlight the best results from the AI. Querying the GPT-3 API (with the Davinci model) can take 30 seconds or more for a large completion, so loading time would have been another problem.
The way it stands now, Random Encounters AI has 6,000+ random encounters in 28 different locations, including arctic, forest, city, village, temple, volcano, and others. The site doesn’t give you every little piece of information that you need, but high-level game masters will enjoy the ideas offered when it comes to random encounters in different locations.
The Future of My Random Encounter Generator
I plan on fine-tuning again. In addition to the original 6,000+ examples, I now have another 6,000 that are available on the Random Encounters AI website. When I get a bit of time (and money), I plan on fine-tuning Davinci again with 1,200 sets of 10 random encounters rather than 600. Hopefully, this will result in even better output and faster AI queries.
I’m also considering using GPT-3 to generate character backstories, spells, locations, quests, and a lot more at LitRPG Adventures. In total, I have over two dozen advanced RPG generators online. We also have a growing library of RPG content that’s already been generated. There’s over 4,000 character backstories, hundreds of monsters, and so much more.
Have Any Questions or Comments?
Do you have questions or comments about my random encounters generator? Feel free to find me on Twitter or email me at email@example.com. I can’t promise I’ll answer, but there’s a good chance I will. Stay tuned as I work on even more projects combining my loves of games, writing, and AI.