Savianto/olmo2-1b-rl-crosscoder

OLMo-2-1B DPO→RLVR1 crosscoder (layer 12, TopK k=64)

A TopK crosscoder trained on paired residual-stream activations of allenai/OLMo-2-0425-1B-DPO and allenai/OLMo-2-0425-1B-RLVR1 to study which features the RL step modifies. Shared encoder, per-model decoders, k=64 active features/token, d_hidden=8192. FVE ≈ 0.87.

Finding: the features RL modifies most are ~2× more reasoning-selective than the ones it leaves alone (corr +0.25), while the full SFT+DPO+RL pipeline points the opposite way — the reasoning- targeting is specific to the RL step. Code + write-up: https://github.com/yash-sawant/rl-crosscoder

from rl_crosscoder.crosscoder import Crosscoder
cc = Crosscoder.from_pretrained("Savianto/olmo2-1b-rl-crosscoder", filename="crosscoder.pt")