cho-ai/constitutional-mt-data
Viewer • Updated • 2.35M • 65 • 1
Data and models for the paper 'Constitutional Midtraining: Content Presence Drives Alignment Gains' .
Note Constitutional midtraining corpus: per-cluster DR/noDR document pools (k1-k4) and assembled training mixes for all five conditions.
Note Shared instruction-tuning mix (200K examples) used for the value-neutral SFT stage.
Note Control, post-MT
Note Control, post-SFT
Note Control, post-BFT
Note Curriculum-DR, post-MT
Note Curriculum-DR, post-SFT
Note Curriculum-DR, post-BFT
Note Uniform-DR, post-MT
Note Uniform-DR, post-SFT
Note Uniform-DR, post-BFT
Note Curriculum-noDR, post-MT
Note Curriculum-noDR, post-SFT
Note Curriculum-noDR, post-BFT
Note Uniform-noDR, post-MT
Note Uniform-noDR, post-SFT
Note Uniform-noDR, post-BFT