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": 74952.30156520358,
-      "required_pages_per_second_upper_estimate": 0.006496675534800273
+      "required_token_positions_per_second": 75012.07721843959,
+      "required_pages_per_second_upper_estimate": 0.006501856736922762
     },
     {
       "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": 9369.037695650448,
-      "required_pages_per_second_upper_estimate": 0.003737449216391594
+      "required_token_positions_per_second": 9376.509652304949,
+      "required_pages_per_second_upper_estimate": 0.0037404298916167817
     },
     {
       "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": 18738.075391300896,
-      "required_pages_per_second_upper_estimate": 0.004334723962913922
+      "required_token_positions_per_second": 18753.019304609898,
+      "required_pages_per_second_upper_estimate": 0.004338180974254068
     },
     {
       "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": 28107.11308695134,
-      "required_pages_per_second_upper_estimate": 0.0017115757203697906
+      "required_token_positions_per_second": 28129.528956914848,
+      "required_pages_per_second_upper_estimate": 0.0017129407292435879
     },
     {
       "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": 28107.11308695134,
-      "required_pages_per_second_upper_estimate": 0.0021637633709023106
+      "required_token_positions_per_second": 28129.528956914848,
+      "required_pages_per_second_upper_estimate": 0.0021654890066232007
     },
     {
       "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": 21,
-      "required_token_positions_per_second": 28107.11308695134,
-      "required_pages_per_second_upper_estimate": 0.34899627608368006
+      "estimated_pages_for_buffer": 22,
+      "required_token_positions_per_second": 28129.528956914848,
+      "required_pages_per_second_upper_estimate": 0.34927460616753603
     }
   ],
   "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": 61,
+    "prefetch_batches": 62,
     "prefetch_capacity_limit": 512,
     "batch_memory_estimate_bytes": 688946,
-    "producer_batch_seconds": 0.15281757075786118,
-    "producer_p95_seconds": 0.2907431249986985,
-    "gpu_update_seconds": 0.9836655907872487,
-    "producer_batches_per_second": 6.543750139730306,
-    "gpu_batches_per_second": 1.0166056527398488,
-    "producer_headroom": 6.4368618471619525,
+    "producer_batch_seconds": 0.12425812846103668,
+    "producer_p95_seconds": 0.3211029350022727,
+    "gpu_update_seconds": 0.9828817269691081,
+    "producer_batches_per_second": 8.047763252072219,
+    "gpu_batches_per_second": 1.0174164119254503,
+    "producer_headroom": 7.909999443435268,
     "sustainable": true,
-    "buffer_seconds": 60.00360103802217,
+    "buffer_seconds": 60.938667072084705,
     "buffer_drain_seconds": null,
     "shard_target_mib": 16,
-    "upload_pages_per_commit": 85,
+    "upload_pages_per_commit": 83,
     "max_pending_pages": 4096,
-    "generation_bytes_per_second": 402755.2997179027,
+    "generation_bytes_per_second": 401658.6071566594,
     "upload_bytes_per_second": 5590558.832680459,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 86.42191165592966,
+    "estimated_full_shards_per_hour": 86.18658696198308,
     "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-v1/1790390190216476996-4d8dc439/train.jsonl

@@ -214,3 +214,4 @@
 {"step": 21400, "loss": 0.6534912511706352, "lr": 0.0003, "grad_norm": 0.13234207093614064, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 3131302397, "tokens_per_second": 148239.98097128168, "data_wait_s": 1.206399974762462e-05, "gpu_peak_gb": 85.43514251708984}
 {"step": 21500, "loss": 0.6984033659100533, "lr": 0.0003, "grad_norm": 0.13349629381036066, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 3146023552, "tokens_per_second": 149947.73587327, "data_wait_s": 1.4861998351989314e-05, "gpu_peak_gb": 85.43514251708984}
 {"step": 21600, "loss": 0.6770138218998909, "lr": 0.0003, "grad_norm": 0.1325722332387709, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 3160724134, "tokens_per_second": 148726.82064154441, "data_wait_s": 1.4473000192083418e-05, "gpu_peak_gb": 85.43514251708984}
+{"step": 21700, "loss": 0.6650346778333187, "lr": 0.0003, "grad_norm": 0.13647737725064946, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 3175255960, "tokens_per_second": 147943.3160290835, "data_wait_s": 1.2603999493876472e-05, "gpu_peak_gb": 85.43514251708984}

runs/void-byte-v1/events.out.tfevents.1790390190.c8125a87-1345-4bfd-b42f-0e32ed8b7bb7-6w295.451.0.1790390190216476996-4d8dc439

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a4c9d034959842962d23c38d85e8fdc4750ba10986f912e6f585bfbc71ca56d3
-size 455011
+oid sha256:23143097a71a3f13f4692745098d478eced025827275846fffdd439d6b68fc00
+size 457129