{ "checkpoint": "llama-3b_layer27_maghead_mlp_h2048x2_t0.05_a1.0.pt", "arch": "maghead_mlp (InfoNCE dir head [MLP 3072->2048->2048->1024 GELU], temp 0.05) + MSE mag head (alpha 1.0)", "model": "meta-llama/Llama-3.2-3B-Instruct", "layer": 27, "data_shards": 24, "val_rank1": 0.624, "val_dir_rank1": 0.638, "val_cos": 0.426, "decode_cos_pred_sbert": 0.397, "decode_cos_upper_sbert": 0.944, "decode_cos_sonar_shortcut": 0.429, "n_decode_eval": 300, "date": "2026-06-06", "note": "Best ALL-AROUND parascope probe. MLP dir head + softer InfoNCE temp (0.05) + higher mag weight (alpha 1.0). vs linear baseline (rank1 0.594, decode 0.363): +5% rank1 AND +9% decode together. Companion rank1-max variant: h2048x2 t0.03 a0.5 (rank1 0.631, decode 0.371)." }