mcgeestocks / boardgamefinds

LLM Content + Boardgames

Project Overview - Board Game Finds

After Llama 2 was open-sourced, I wanted to see what type of novel content I could create and how that would compare against OpenAI's offering. Of the 2500+ pages, half of the content on boardgamefinds.com is generated by OpenAI's GPT-3.5 Turbo, and the other half is generated by Llama 2. While getting to "quality" outputs with Llama took more trial and error, the content is nearly indistinguishable from that of OpenAI. Initially, I intended this to be a small project. However, if I were to invest more time, I'd probably start with a VectorDB and LangChain, as well as trying to do in-browser recommendations using ONNX Web Runtime

Build with:

• NextJS 13 + Typescript
• TailwindCSS
• MongoDB Atlas
• Running on Vercel/Edge

Content Generation

Generating everything took place over three steps. First, I compiled a dataset of board games and their relationships using sites like Board Game Atlas, Board Game Geek, and others.

Next, I converted these relationships into JSON format and prepared a template. Finally, I fed them into Guidance or Llama.Cpp to generate the markdown. This markdown is served from MongoDB Atlas and converted to HTML on-the-fly using react-markdown.

The result is a site with 2500+ pages of content, all generated by either GPT-3.5 Turbo or Llama 2.

Step 1: Data Sample

{ "Hand Management": ["Terraforming Mars", "Brass: Birmingham", "Ark Nova"] }

Step 2: Template Creation

{{#system~}}
  Your are a an expert Game Designer and Copy Writer.
{{~/system}}

{{#user~}}
  Write a short entry for the Board Game Mechanic Encyclopedia on "{{mechanic}}",
  using
  {{n_1}},
  {{n_2}},
  {{n_3}}
  as examples to compare and contrast how the mechanic effects gameplay and why
  people enjoy it. The entry must be concise and comprehensive overview with a #
  Overview, # Gameplay/Functionality, # Examples from Popular Games sections. #
  Overview
{{~/user}}

Step 3: See the Result