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": 76319.31028689623,
-      "required_pages_per_second_upper_estimate": 0.006615164386144662
+      "required_token_positions_per_second": 76075.13267074381,
+      "required_pages_per_second_upper_estimate": 0.00659399968923908
     },
     {
       "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": 9539.913785862029,
-      "required_pages_per_second_upper_estimate": 0.0038056142436022136
+      "required_token_positions_per_second": 9509.391583842977,
+      "required_pages_per_second_upper_estimate": 0.0037934384808692265
     },
     {
       "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": 19079.827571724058,
-      "required_pages_per_second_upper_estimate": 0.00441378231522187
+      "required_token_positions_per_second": 19018.783167685953,
+      "required_pages_per_second_upper_estimate": 0.004399660766692472
     },
     {
       "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": 28619.74135758609,
-      "required_pages_per_second_upper_estimate": 0.0017427920924987705
+      "required_token_positions_per_second": 28528.17475152893,
+      "required_pages_per_second_upper_estimate": 0.0017372161666027433
     },
     {
       "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": 28619.74135758609,
-      "required_pages_per_second_upper_estimate": 0.0022032269142216493
+      "required_token_positions_per_second": 28528.17475152893,
+      "required_pages_per_second_upper_estimate": 0.002196177863414781
     },
     {
       "source": "cortex",
@@ -92,8 +92,8 @@
       "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": 28619.74135758609,
-      "required_pages_per_second_upper_estimate": 0.35536140354850676
+      "required_token_positions_per_second": 28528.17475152893,
+      "required_pages_per_second_upper_estimate": 0.35422445275499376
     }
   ],
   "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": 63,
+    "prefetch_batches": 62,
     "prefetch_capacity_limit": 512,
     "batch_memory_estimate_bytes": 690670,
-    "producer_batch_seconds": 0.0642758016720677,
-    "producer_p95_seconds": 0.21357675599938375,
-    "gpu_update_seconds": 0.9660464661282303,
-    "producer_batches_per_second": 15.557954533215405,
-    "gpu_batches_per_second": 1.0351468951673208,
-    "producer_headroom": 15.029706996996422,
+    "producer_batch_seconds": 0.14040043821105996,
+    "producer_p95_seconds": 0.28223465400151326,
+    "gpu_update_seconds": 0.9691471761094088,
+    "producer_batches_per_second": 7.122484892082236,
+    "gpu_batches_per_second": 1.0318350242885175,
+    "producer_headroom": 6.902736120043426,
     "sustainable": true,
-    "buffer_seconds": 60.860927366078506,
+    "buffer_seconds": 60.087124918783346,
     "buffer_drain_seconds": null,
     "shard_target_mib": 16,
-    "upload_pages_per_commit": 80,
+    "upload_pages_per_commit": 83,
     "max_pending_pages": 4096,
-    "generation_bytes_per_second": 415744.03981670085,
+    "generation_bytes_per_second": 430417.29008097673,
     "upload_bytes_per_second": 3972863.3572530756,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 89.20899291873712,
+    "estimated_full_shards_per_hour": 92.35753084966638,
     "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-v1/1790703713325746110-adad6d18/train.jsonl

@@ -112,3 +112,4 @@
 {"step": 291400, "loss": 0.5936985500156879, "lr": 0.0003, "grad_norm": 0.16376809380638718, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 42652512807, "tokens_per_second": 150215.8051876788, "data_wait_s": 1.1799002095358446e-05, "gpu_peak_gb": 85.46991729736328}
 {"step": 291500, "loss": 0.565649164840579, "lr": 0.0003, "grad_norm": 0.1470336876474914, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 42667207381, "tokens_per_second": 150758.1432170335, "data_wait_s": 8.61400258145295e-06, "gpu_peak_gb": 85.46991729736328}
 {"step": 291600, "loss": 0.6152579374611378, "lr": 0.0003, "grad_norm": 0.18157318205932949, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 42681906131, "tokens_per_second": 150530.93471218826, "data_wait_s": 9.913997928379104e-06, "gpu_peak_gb": 85.46991729736328}
+{"step": 291700, "loss": 0.5707413833588362, "lr": 0.0003, "grad_norm": 0.14450361581721619, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 42696551761, "tokens_per_second": 149938.74250766585, "data_wait_s": 1.3040000339969993e-05, "gpu_peak_gb": 85.46991729736328}

runs/void-byte-v1/events.out.tfevents.1790703713.3e6d470a-8cc7-4d7e-a224-51a0f9a2ce55-85btc.38629.0.1790703713325746110-adad6d18

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:557e60b6b58c13b0c8b3ca6fdeabac7adf51d0022017450fa748b1499e08e517
-size 243183
+oid sha256:d244c495019f99fef1e3ee25bd09a6bfa2d93e8b7ff7425210fd2f41127376c3
+size 245301