Pinned Loading

1. dataset_foundry dataset_foundry Public

   A toolkit for building validated datasets. Uses the concept of data pipelines to load, generate and validate datasets, especially for those used in AI safety evaluations.

   Python 1

2. single_file_backdoors single_file_backdoors Public

   Evaluates how AI models might inject backdoors when refactoring single files and how to detect and defend against such insertions.

   Python 1 1

3. full-repo-refactor full-repo-refactor Public

   A Control Arena setting for evaluating agents inserting backdoors while refactoring full repos.

   Python

4. full_repo_datasets full_repo_datasets Public

   Contains datasets of full repos for doing AI safety research, along with Dataset Foundry pipelines for generating repos and datasets.

   Python

5. llm-action-evals llm-action-evals Public

   A framework that allows non-programmers to build AI safety evals of LLMs taking actions in the real world via function-calling.

   Python

6. ai-digest-demo ai-digest-demo Public

   Demo of targeted AI persuasion by building a profile of user using Facebook Likes. Developed as an 8-hour take-home test for AI Digest.

   TypeScript