---
base_model: geodesic-research/nemotron-super-120b-cc-mt-replay-only-1b-200k-sft
tags: [nemotron, task-vector, constitutional-midtraining, graft]
---
# Task-vector graft: curriculum-DR constitutional midtraining applied post-SFT

`merged = geodesic-research/nemotron-super-120b-cc-mt-replay-only-1b-200k-sft + 1.0 * (geodesic-research/nemotron-super-120b-cc-mt-curriculum-dr - geodesic-research/nemotron-super-120b-cc-mt-replay-only-1b)`

All three source checkpoints are from the Constitutional Midtraining paper
(arXiv:2607.26654, geodesic-research). This model tests whether the
constitutional-midtraining weight delta (taken post-midtraining, replay
cancelled via the control anchor) installs the same alignment behaviour when
added to the control *after* its 200K SFT — i.e., whether the effect is
portable content rather than midtraining placement. Built shard-wise in fp32,
cast back to bf16; exact source revisions and delta norms in
`task_vector_report.json`.