add merge_weights.json

merge_weights.json

@@ -0,0 +1,72 @@
+{
+  "merged": "/home/elie/anthropic-research-jspace/runs/E7_inkling/fits/Inkling-NVFP4_jacobian_lens.pt",
+  "merge_rule": "exact n_prompts-weighted mean: J = (n0*J0 + n1*J1)/(n0+n1); code/merge_lens.py -> jlens.lens.JacobianLens.merge",
+  "shards": {
+    "shard0": {
+      "path": "/home/elie/anthropic-research-jspace/runs/E7_inkling/fits/shard0/Inkling-NVFP4_shard0of2_jacobian_lens.pt",
+      "n_prompts": 125,
+      "weight": 0.5
+    },
+    "shard1": {
+      "path": "/home/elie/anthropic-research-jspace/runs/E7_inkling/fits/shard1/Inkling-NVFP4_shard1of2_jacobian_lens.pt",
+      "n_prompts": 125,
+      "weight": 0.5
+    }
+  },
+  "n_prompts_merged": 250,
+  "storage_dtype": "float32 (shards are fp32; fp16 default overridden so the merged artifact is the exact mean)",
+  "n_valid_positions_per_prompt": 111,
+  "note": "every prompt in both shards has seq_len=128 / n_valid=111, so n_prompts-weighting == n_valid/pair-count weighting exactly; weights 0.5/0.5",
+  "prompt_partition": "global wikitext prompts [0..249]; shard0 = prompts[0::2] (even), shard1 = prompts[1::2] (odd) via fit_moe.py --prompt_offset/--prompt_stride",
+  "sanity_spot_checks_norm_over_sqrt_d": {
+    "0": {
+      "norm_shard0": 200.0007228777548,
+      "norm_shard1": 189.64136021348878,
+      "norm_merged": 169.12728229765082,
+      "max_abs_dev_from_elementwise_mean": 3.0517578125e-05,
+      "norm_le_mean_of_shard_norms": true
+    },
+    "1": {
+      "norm_shard0": 174.78741660064878,
+      "norm_shard1": 165.39693347753584,
+      "norm_merged": 145.8192053089544,
+      "max_abs_dev_from_elementwise_mean": 1.52587890625e-05,
+      "norm_le_mean_of_shard_norms": true
+    },
+    "2": {
+      "norm_shard0": 91.02883490107857,
+      "norm_shard1": 84.2912555172821,
+      "norm_merged": 71.81201800576076,
+      "max_abs_dev_from_elementwise_mean": 7.62939453125e-06,
+      "norm_le_mean_of_shard_norms": true
+    },
+    "16": {
+      "norm_shard0": 31.045034931439393,
+      "norm_shard1": 30.615423000958657,
+      "norm_merged": 27.741454332323375,
+      "max_abs_dev_from_elementwise_mean": 1.52587890625e-05,
+      "norm_le_mean_of_shard_norms": true
+    },
+    "32": {
+      "norm_shard0": 1.8632074850435494,
+      "norm_shard1": 1.8687422843306152,
+      "norm_merged": 1.8487818109309748,
+      "max_abs_dev_from_elementwise_mean": 4.76837158203125e-07,
+      "norm_le_mean_of_shard_norms": true
+    },
+    "48": {
+      "norm_shard0": 1.1597017330407198,
+      "norm_shard1": 1.1585713935619513,
+      "norm_merged": 1.158697635730552,
+      "max_abs_dev_from_elementwise_mean": 4.76837158203125e-07,
+      "norm_le_mean_of_shard_norms": true
+    },
+    "64": {
+      "norm_shard0": 1.0228682650302547,
+      "norm_shard1": 1.0229844817991904,
+      "norm_merged": 1.0229251080730173,
+      "max_abs_dev_from_elementwise_mean": 1.1920928955078125e-07,
+      "norm_le_mean_of_shard_norms": true
+    }
+  }
+}