Upload gemma-27b_layer58_ml50-54-58_maghead_mlp_h2048x2_t0.05_a1.0.metrics.json with huggingface_hub

gemma-27b_layer58_ml50-54-58_maghead_mlp_h2048x2_t0.05_a1.0.metrics.json

@@ -0,0 +1,19 @@
+{
+  "checkpoint": "gemma-27b_layer58_ml50-54-58_maghead_mlp_h2048x2_t0.05_a1.0.pt",
+  "arch": "maghead_mlp, MULTI-LAYER input (concat residuals of layers 50,54,58 -> 16128-d) + linear mag head",
+  "base_model": "gemma-3-27b",
+  "data": "fineweb-gemma27b-residuals (24 shards)",
+  "layers": [
+    50,
+    54,
+    58
+  ],
+  "d_res": 16128,
+  "data_shards": 24,
+  "val_rank1": 0.676,
+  "val_dir_rank1": 0.693,
+  "val_cos": 0.443,
+  "val_mag_mse": 0.685,
+  "date": "2026-06-07",
+  "note": "OVERALL RETRIEVAL SOTA across all parascope probes. Multi-layer input (3 mid+deep layers) + 24sh data stack: single-L58-16sh 0.658 -> ml3-12sh 0.661 -> ml3-24sh 0.676. dir_rank1 0.693. Decode still magnitude-capped (mag_mse ~0.69); multi-layer lifts direction/retrieval, not the magnitude wall."
+}