ProgressGym: Alignment with a Millennium of Moral Progress
Paper
•
2406.20087
•
Published
•
3
Alignment with a millennium of moral progress
Note Leaderboard + interactive playground. Some functionalities are currently under construction.
Note The central dataset containing 9 centuries of historical text data. Used in the training of HistLlama models.
Note Timeless and value-neutral instruction-tuning data. Used in the training of HistLlama models.
Note Demonstrative dataset containing prompts and response options in the morality evaluation pipeline. Used when benchmarking algorithms against the ProgressGym challenges.
Note Historical LLM (HistLlama) tuned on the corresponding century's text data. Used when benchmarking algorithms against the ProgressGym challenges. Likewise for the other 35 historical LLMs.