Upload gemma-27b_layer58_maghead_mlp_h2048x2_t0.05_a2.0.metrics.json with huggingface_hub

gemma-27b_layer58_maghead_mlp_h2048x2_t0.05_a2.0.metrics.json

@@ -0,0 +1,18 @@
+{
+  "checkpoint": "gemma-27b_layer58_maghead_mlp_h2048x2_t0.05_a2.0.pt",
+  "arch": "maghead_mlp (MLP dir head 5376->2048->2048->1024 GELU + linear mag head)",
+  "base_model": "gemma-3-27b",
+  "data": "nickypro/fineweb-gemma27b-residuals (16 shards)",
+  "layer": 58,
+  "d_res": 5376,
+  "data_shards": 16,
+  "val_rank1": 0.658,
+  "val_dir_rank1": 0.676,
+  "val_cos": 0.43,
+  "val_mag_mse": 0.7,
+  "decode_cos_pred_sbert": 0.343,
+  "decode_cos_upper_sbert": 0.801,
+  "decode_cos_sonar_shortcut": 0.431,
+  "date": "2026-06-06",
+  "note": "RETRIEVAL SOTA across all parascope probes: rank1 0.658 / dir_rank1 0.676 (> llama-3b champion 0.624/0.638) with only 16 shards. Bigger base model -> better DIRECTION prediction. Magnitude head is the wall (mag_mse 0.70); decode ~0.43 of this corpus's roundtrip ceiling 0.80 (same fraction as llama-3b's 0.42). Different corpus (fineweb) than the llama-3b probes, so absolute decode not directly comparable."
+}