head Doctor rank ladder: held-out KL + probe vs branch size

head Doctor rank 512 on the FP-body-trained ternary head: KL 0.0470->0.0412, probe 1.0740x -> 1.0669x

head Doctor rank 256 on the FP-body-trained ternary head: KL 0.0470->0.0428, probe 1.0740x -> 1.0683x

head Doctor rank 64 on the FP-body-trained ternary head: KL 0.0470->0.0429, probe 1.0740x -> 1.0691x

head Doctor rank 16 on the FP-body-trained ternary head: KL 0.0470->0.0451, probe 1.0740x -> 1.0712x

head Doctor rank 64 on the FP-body-trained ternary head: KL 0.0470->0.0449, probe 1.0748x -> 1.0719x

shadow head on the FULL FP body v2 (lloyd 3, cosine, batch 4->8, polyak 200, init shadow_head.safetensors): held-out KL 0.0788->0.0470, wikitext probe 1.0740x, 1500 steps

shadow head trained on the FULL FP body (head gaining everything): KL 0.1844->0.0788, head probe 1.1028x, 1500 steps

re-fit k1-KV branches (v2, trained with live Doctors)

MERGED Doctors + re-fit k1-KV rank-16 (LIVE-Doctors training; held-out gap +0.0874 -> -0.0043, 105% closed)

trained k1-KV compensation branches (rank 16, k/v_proj, post-RoPE rotation): held-out gap +0.1711 -> -0.0179

Upload folder using huggingface_hub

initial commit