Update TensorBoard training metrics

benchmark_cache/shard_scheduler.json

@@ -22,8 +22,8 @@
       "tokens_per_second": 1062967.6168812835,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 1,
-      "required_token_positions_per_second": 75761.30885298371,
-      "required_pages_per_second_upper_estimate": 0.006566798236094829
+      "required_token_positions_per_second": 74374.54380643701,
+      "required_pages_per_second_upper_estimate": 0.006446596956584038
     },
     {
       "source": "rewrite6",
@@ -36,8 +36,8 @@
       "tokens_per_second": 1206672.263377472,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 1,
-      "required_token_positions_per_second": 9470.163606622964,
-      "required_pages_per_second_upper_estimate": 0.0037777898542456375
+      "required_token_positions_per_second": 9296.817975804626,
+      "required_pages_per_second_upper_estimate": 0.0037086396903640603
     },
     {
       "source": "openmath",
@@ -50,8 +50,8 @@
       "tokens_per_second": 410042.7581134458,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 1,
-      "required_token_positions_per_second": 18940.327213245928,
-      "required_pages_per_second_upper_estimate": 0.004381511362410412
+      "required_token_positions_per_second": 18593.635951609253,
+      "required_pages_per_second_upper_estimate": 0.004301310440588577
     },
     {
       "source": "ultra_qa",
@@ -64,8 +64,8 @@
       "tokens_per_second": 2163486.5624068812,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 1,
-      "required_token_positions_per_second": 28410.490819868894,
-      "required_pages_per_second_upper_estimate": 0.001730049832604518
+      "required_token_positions_per_second": 27890.453927413877,
+      "required_pages_per_second_upper_estimate": 0.0016983823142767188
     },
     {
       "source": "ultra_style",
@@ -78,8 +78,8 @@
       "tokens_per_second": 1546694.2737633225,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 1,
-      "required_token_positions_per_second": 28410.490819868894,
-      "required_pages_per_second_upper_estimate": 0.0021871182285856184
+      "required_token_positions_per_second": 27890.453927413877,
+      "required_pages_per_second_upper_estimate": 0.0021470843490501773
     },
     {
       "source": "cortex",
@@ -91,9 +91,9 @@
       "read_seconds": 0.042780299000014566,
       "tokens_per_second": 1015177.8395898147,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
-      "estimated_pages_for_buffer": 22,
-      "required_token_positions_per_second": 28410.490819868894,
-      "required_pages_per_second_upper_estimate": 0.35276321218655887
+      "estimated_pages_for_buffer": 21,
+      "required_token_positions_per_second": 27890.453927413877,
+      "required_pages_per_second_upper_estimate": 0.3463060944337867
     }
   ],
   "gpu_measurement": "Selected forward/backward estimate; replaced by live successful update timings.",
@@ -155,27 +155,27 @@
   "shard_benchmark_scope": "Local packing/read measurements; upload capacity measured only on real data commits. Targets larger than available pages are sample-limited.",
   "estimated_fresh_generation_batch_seconds": 0.1434574083060457,
   "selected": {
-    "prefetch_batches": 62,
+    "prefetch_batches": 61,
     "prefetch_capacity_limit": 512,
     "batch_memory_estimate_bytes": 690670,
-    "producer_batch_seconds": 0.6132215871484448,
-    "producer_p95_seconds": 0.35302726599911693,
-    "gpu_update_seconds": 0.9731616456504284,
-    "producer_batches_per_second": 1.6307318935886161,
-    "gpu_batches_per_second": 1.0275785163436375,
-    "producer_headroom": 1.586965733179337,
+    "producer_batch_seconds": 0.25133367493759806,
+    "producer_p95_seconds": 0.35121369899934507,
+    "gpu_update_seconds": 0.9913069207104024,
+    "producer_batches_per_second": 3.978774433025273,
+    "gpu_batches_per_second": 1.008769311610745,
+    "producer_headroom": 3.9441866314035607,
     "sustainable": true,
-    "buffer_seconds": 60.33602203032656,
+    "buffer_seconds": 60.469722163334545,
     "buffer_drain_seconds": null,
     "shard_target_mib": 16,
-    "upload_pages_per_commit": 82,
+    "upload_pages_per_commit": 80,
     "max_pending_pages": 4096,
-    "generation_bytes_per_second": 371255.4969490323,
+    "generation_bytes_per_second": 359864.87175570894,
     "upload_bytes_per_second": 4904741.672189561,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 79.66278725960947,
+    "estimated_full_shards_per_hour": 77.21862425330592,
     "selection_policy": "Smallest shard within 90% of measured local peak, enlarged to amortize 30 seconds of measured page production; bounded to 256 MiB."
   },
   "upload_measurements": [

metrics/void-byte-v1/1790651964532391049-1e9227c4/train.jsonl

@@ -315,3 +315,4 @@
 {"step": 268600, "loss": 0.5550219640135765, "lr": 0.0003, "grad_norm": 0.1480452006594306, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 39314917054, "tokens_per_second": 149776.29686238288, "data_wait_s": 1.7141002899734303e-05, "gpu_peak_gb": 85.46427917480469}
 {"step": 268700, "loss": 0.5741091929376125, "lr": 0.0003, "grad_norm": 0.1482366127425559, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 39329559086, "tokens_per_second": 149792.38781241022, "data_wait_s": 1.2163000064902008e-05, "gpu_peak_gb": 85.46427917480469}
 {"step": 268800, "loss": 0.5637525208294392, "lr": 0.0003, "grad_norm": 0.15044971965244122, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 39344183795, "tokens_per_second": 134925.3938784452, "data_wait_s": 1.3763998140348122e-05, "gpu_peak_gb": 85.46427917480469}
+{"step": 268900, "loss": 0.5969801694154739, "lr": 0.0003, "grad_norm": 0.14266062221497894, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 39358934556, "tokens_per_second": 150935.2554989693, "data_wait_s": 1.1551001080079004e-05, "gpu_peak_gb": 85.46427917480469}

runs/void-byte-v1/events.out.tfevents.1790651964.96f3217e-af24-484e-a73b-abea9eb37bd2-c4mbg.442.0.1790651964532391049-1e9227c4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6e8f09c6aafb7225104513dd01907d55e79329de6b09be5dc04e512360066306
-size 676269
+oid sha256:e95773ff8e6f260d66bcf7d5ccdc864416281e4db863684a1a186637c4be2cdb
+size 678387