Skip to content

Repository files navigation

mafia

Benchmark models with adversarial games such as Mafia.

Intro

Existing LM benchmarks in the format of pre-written input prompts are static and thus easy to optimize for. In this context growth from hillclimbing does not necessarily correlate with actual progress towards AGI, which makes benchmarking ineffective.

On the other hand, adversarial games provide an attractive alternative approach to benchmarking. In order to win at Mafia, a model needs to engage in multi-turn dialogue, infer hidden intentions, and provide false information in context. These abilities are observed from a dynamic environment, which makes it difficult to optimize over.

Deception as a modeling ability has attracted interest since the Turing Test. Which model will win the ultimate test of intelligence?

Run code

Prepare .env file locally in this format:

OPENAI_API_KEY=sk-proj-xxx_abc

Then start Docker:

docker-compose up

Contributing

  • Format your code with black before submitting. This helps by producing the smallest possible diff.
  • Commit messages should start with tags. E.g. [model] add Anthropic provider

About

No description, website, or topics provided.

Resources

Stars

5 stars

Watchers

2 watching

Forks

Releases

Packages

Used by

Contributors

Languages