Upload results.json with huggingface_hub

results.json

@@ -19,6 +19,13 @@
       "val": 2.360668977101644,
       "ppl": 10.598038926854809,
       "params": 2104448
+    },
+    "qat_ternary_embed": {
+      "val": 2.310712154706319,
+      "ppl": 10.081601759246057,
+      "params": 11159360,
+      "embed_mode": "ternary",
+      "qat_recovery_of_ptq_embed_loss": 0.760477491978586
     }
   },
   "baselines": {
@@ -72,5 +79,45 @@
     "weight_zero_fraction": 0.314,
     "qat_wallclock_multiplier": 1.1170731707317074
   },
-  "note": "checkpoints store fp32 MASTER weights. ternary quantization happens in the forward pass. the 4.61 MB figure is the packed inference size."
+  "note": "checkpoints store fp32 MASTER weights. ternary quantization happens in the forward pass. the 4.61 MB figure is the packed inference size.",
+  "pareto": {
+    "all_fp16": {
+      "bytes": 22318720,
+      "MB": 22.32,
+      "share_1p58bit": 0.0,
+      "share_fp16": 0.0016,
+      "compression_vs_fp16": 1.0,
+      "val": 2.0553244292736053
+    },
+    "ternary_body_fp16_embed": {
+      "bytes": 4605521,
+      "MB": 4.61,
+      "share_1p58bit": 0.4229,
+      "share_fp16": 0.0079,
+      "compression_vs_fp16": 4.85,
+      "val": 2.1759746789932253
+    },
+    "ternary_body_int8_embed": {
+      "bytes": 3294801,
+      "MB": 3.29,
+      "share_1p58bit": 0.5911,
+      "share_fp16": 0.0111,
+      "compression_vs_fp16": 6.77,
+      "val": 2.1795
+    },
+    "ternary_body_ternary_embed": {
+      "bytes": 2243762,
+      "MB": 2.24,
+      "share_1p58bit": 0.9837,
+      "share_fp16": 0.0163,
+      "compression_vs_fp16": 9.95,
+      "val": 2.310712154706319
+    }
+  },
+  "param_buckets": {
+    "ternary_body": 9830400,
+    "embedding_tied": 1310720,
+    "norms_always_fp16": 18240
+  },
+  "correction": "an earlier table bucketed embedding+norms as 'full precision' then quantized the bucket; the true fp16 share of the ternary-embed model is 1.63%, not 11.9%"
 }