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": 74435.11070962898,
-      "required_pages_per_second_upper_estimate": 0.006451846742247306
+      "required_token_positions_per_second": 73592.03837535501,
+      "required_pages_per_second_upper_estimate": 0.006378771369059734
     },
     {
       "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": 9304.388838703622,
-      "required_pages_per_second_upper_estimate": 0.0037116598207689575
+      "required_token_positions_per_second": 9199.004796919377,
+      "required_pages_per_second_upper_estimate": 0.003669620550869386
     },
     {
       "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": 18608.777677407244,
-      "required_pages_per_second_upper_estimate": 0.0043048132123666705
+      "required_token_positions_per_second": 18398.009593838753,
+      "required_pages_per_second_upper_estimate": 0.004256055725624679
     },
     {
       "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": 27913.166516110865,
-      "required_pages_per_second_upper_estimate": 0.0016997653917646218
+      "required_token_positions_per_second": 27597.01439075813,
+      "required_pages_per_second_upper_estimate": 0.0016805133860526493
     },
     {
       "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": 27913.166516110865,
-      "required_pages_per_second_upper_estimate": 0.002148832827000541
+      "required_token_positions_per_second": 27597.01439075813,
+      "required_pages_per_second_upper_estimate": 0.0021244945612257913
     },
     {
       "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": 21,
-      "required_token_positions_per_second": 27913.166516110865,
-      "required_pages_per_second_upper_estimate": 0.34658810877125873
+      "required_token_positions_per_second": 27597.01439075813,
+      "required_pages_per_second_upper_estimate": 0.3426625574674762
     }
   ],
   "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": 60,
     "prefetch_capacity_limit": 512,
     "batch_memory_estimate_bytes": 688946,
-    "producer_batch_seconds": 0.5052233389531295,
-    "producer_p95_seconds": 0.37241947000075015,
-    "gpu_update_seconds": 0.9905003068728223,
-    "producer_batches_per_second": 1.97932265376357,
-    "gpu_batches_per_second": 1.0095908028107228,
-    "producer_headroom": 1.9605196959531452,
+    "producer_batch_seconds": 0.1376259003438065,
+    "producer_p95_seconds": 0.3576471579981444,
+    "gpu_update_seconds": 1.00184750453509,
+    "producer_batches_per_second": 7.266074172825583,
+    "gpu_batches_per_second": 0.9981559024435087,
+    "producer_headroom": 7.27949827781218,
     "sustainable": true,
-    "buffer_seconds": 60.42051871924216,
+    "buffer_seconds": 60.1108502721054,
     "buffer_drain_seconds": null,
     "shard_target_mib": 16,
-    "upload_pages_per_commit": 79,
+    "upload_pages_per_commit": 75,
     "max_pending_pages": 4096,
-    "generation_bytes_per_second": 268212.57396110793,
+    "generation_bytes_per_second": 279407.0488489787,
     "upload_bytes_per_second": 3927731.034962253,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 57.552174702882084,
+    "estimated_full_shards_per_hour": 59.95424841978093,
     "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-v1/1790519810984249790-cb9c9513/train.jsonl

@@ -276,3 +276,4 @@
 {"step": 142600, "loss": 0.5838275533169508, "lr": 0.0003, "grad_norm": 0.14670436260470357, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 20869587389, "tokens_per_second": 149532.45504962013, "data_wait_s": 1.2805001460947096e-05, "gpu_peak_gb": 85.46427917480469}
 {"step": 142700, "loss": 0.6139459684491158, "lr": 0.0003, "grad_norm": 0.13226853062715044, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 20884228808, "tokens_per_second": 148261.82778590324, "data_wait_s": 1.6253998182946816e-05, "gpu_peak_gb": 85.46427917480469}
 {"step": 142800, "loss": 0.5808074194937944, "lr": 0.0003, "grad_norm": 0.15253148407926956, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 20898723309, "tokens_per_second": 147789.7326634042, "data_wait_s": 9.75099828792736e-06, "gpu_peak_gb": 85.46427917480469}
+{"step": 142900, "loss": 0.6225517578423023, "lr": 0.0003, "grad_norm": 0.13541117666423091, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 20913218827, "tokens_per_second": 148693.3120301591, "data_wait_s": 1.123200127040036e-05, "gpu_peak_gb": 85.46427917480469}

runs/void-byte-v1/events.out.tfevents.1790519810.cad0b19a-59f2-43f0-8fb4-2a94097843e2-tjh7c.1336.0.1790519810984249790-cb9c9513

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0e0cbddb5b018ec0bf2537d9ba425f7288a35dccb48a9e625774b8ac03bb988e
-size 592884
+oid sha256:7b242176661248e3ef307f6ef8aa92429f4702ef6c10a4a295a345c161d15ad7
+size 595002