Note Labeled test set for whether data is not related to AI, neutral AI discourse, AI misalignment, or positive AI discourse

Note LessWrong and documents related to AI alignment

Note Filtered and reformatted version of Anthropic's propensity evaluations

Note Model organisms dataset made of of both LessWrong and general data

Note Model organisms dataset made of of just general data

Note Tulu-3 generic instruction following datasets. Used string matching to remove most refusals or discussions of AI

Note A sample of documents from DCLM that reference AI science fictions

Kyle1668/pt_alignment_continue_baseline_v1_7

Text Generation • 7B • Updated Oct 5 • 80

Note Continual pretraining on LessWrong: Seed=1234

Kyle1668/pt_alignment_continue_baseline_v1_7_seed_1

Text Generation • 7B • Updated Oct 6 • 6

Note Continual pretraining on LessWrong: Seed=1

Kyle1668/pt_alignment_continue_baseline_v1_7_seed_42

Text Generation • 7B • Updated Oct 6 • 8

Note Continual pretraining on LessWrong: Seed=42

Kyle1668/pt_alignment_continue_baseline_v1_7_replay_only

Text Generation • 7B • Updated Oct 5 • 10

Note Continual pretraining on replay data unrelated to AI: Seed=1234

Kyle1668/pt_alignment_continue_baseline_v1_7_replay_only_seed_1

Text Generation • 7B • Updated Oct 6 • 6

Note Continual pretraining on replay data unrelated to AI: Seed=1

Kyle1668/pt_alignment_continue_baseline_v1_7_replay_only_seed_42

Text Generation • 7B • Updated Oct 6 • 8

Note Continual pretraining on replay data unrelated to AI: Seed=42