Update TensorBoard training metrics

benchmark_cache/shard_scheduler.json

@@ -22,8 +22,8 @@
       "tokens_per_second": 598315.063954628,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 1,
-      "required_token_positions_per_second": 77453.10452099967,
-      "required_pages_per_second_upper_estimate": 0.005358385486725273
+      "required_token_positions_per_second": 77295.74694209894,
+      "required_pages_per_second_upper_estimate": 0.005347499124297034
     },
     {
       "source": "rewrite6",
@@ -36,8 +36,8 @@
       "tokens_per_second": 8533.957213384612,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 1,
-      "required_token_positions_per_second": 9681.638065124958,
-      "required_pages_per_second_upper_estimate": 0.004053499682274867
+      "required_token_positions_per_second": 9661.968367762367,
+      "required_pages_per_second_upper_estimate": 0.004045264390739139
     },
     {
       "source": "openmath",
@@ -50,8 +50,8 @@
       "tokens_per_second": 380144.35761751566,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 1,
-      "required_token_positions_per_second": 19363.276130249917,
-      "required_pages_per_second_upper_estimate": 0.004032859774000186
+      "required_token_positions_per_second": 19323.936735524734,
+      "required_pages_per_second_upper_estimate": 0.004024666415528535
     },
     {
       "source": "ultra_qa",
@@ -64,8 +64,8 @@
       "tokens_per_second": 43742.16556775435,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 1,
-      "required_token_positions_per_second": 29044.914195374873,
-      "required_pages_per_second_upper_estimate": 0.0017411982569966386
+      "required_token_positions_per_second": 28985.905103287103,
+      "required_pages_per_second_upper_estimate": 0.0017376607520276433
     },
     {
       "source": "ultra_style",
@@ -78,8 +78,8 @@
       "tokens_per_second": 22018.86276523721,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 1,
-      "required_token_positions_per_second": 29044.914195374873,
-      "required_pages_per_second_upper_estimate": 0.0022581143236144238
+      "required_token_positions_per_second": 28985.905103287103,
+      "required_pages_per_second_upper_estimate": 0.002253526626258165
     },
     {
       "source": "cortex",
@@ -92,8 +92,8 @@
       "tokens_per_second": 910315.184299418,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 23,
-      "required_token_positions_per_second": 29044.914195374873,
-      "required_pages_per_second_upper_estimate": 0.37255697329914794
+      "required_token_positions_per_second": 28985.905103287103,
+      "required_pages_per_second_upper_estimate": 0.37180006802487275
     }
   ],
   "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": 2.558590814491007,
   "selected": {
-    "prefetch_batches": 64,
+    "prefetch_batches": 63,
     "prefetch_capacity_limit": 512,
     "batch_memory_estimate_bytes": 720750,
-    "producer_batch_seconds": 0.43779749998429907,
-    "producer_p95_seconds": 0.6190619110057014,
-    "gpu_update_seconds": 0.9519050328061455,
-    "producer_batches_per_second": 2.2841610562779904,
-    "gpu_batches_per_second": 1.0505249636637324,
-    "producer_headroom": 2.1743044052108202,
+    "producer_batch_seconds": 0.168329123476326,
+    "producer_p95_seconds": 0.6096285350067774,
+    "gpu_update_seconds": 0.9538429074917734,
+    "producer_batches_per_second": 5.940742631744537,
+    "gpu_batches_per_second": 1.0483906649047707,
+    "producer_headroom": 5.666535224523538,
     "sustainable": true,
-    "buffer_seconds": 60.92192209959331,
+    "buffer_seconds": 60.09210317198172,
     "buffer_drain_seconds": null,
     "shard_target_mib": 16,
-    "upload_pages_per_commit": 80,
+    "upload_pages_per_commit": 82,
     "max_pending_pages": 4096,
-    "generation_bytes_per_second": 348183.8477276332,
+    "generation_bytes_per_second": 333394.84087974246,
     "upload_bytes_per_second": 4993838.710405627,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 74.7121484172034,
+    "estimated_full_shards_per_hour": 71.53877181810574,
     "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-v1/1791463982175382982-d4a3b372/train.jsonl

@@ -399,3 +399,4 @@
 {"step": 964900, "loss": 0.5310418605804443, "lr": 3e-05, "grad_norm": 0.3544124334635774, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 141234854891, "tokens_per_second": 151755.32677238516, "data_wait_s": 9.47499938774854e-06, "gpu_peak_gb": 89.37899398803711}
 {"step": 965000, "loss": 0.5895737409591675, "lr": 3e-05, "grad_norm": 0.3801122804407297, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 141249570216, "tokens_per_second": 153698.0028411782, "data_wait_s": 1.350299862679094e-05, "gpu_peak_gb": 89.37899398803711}
 {"step": 965100, "loss": 0.5409970879554749, "lr": 3e-05, "grad_norm": 0.34271652718062795, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 141264249572, "tokens_per_second": 131399.95456990355, "data_wait_s": 1.9553997844923288e-05, "gpu_peak_gb": 89.37899398803711}
+{"step": 965200, "loss": 0.5364992022514343, "lr": 3e-05, "grad_norm": 0.3412135978278283, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 141279198323, "tokens_per_second": 154604.31451173962, "data_wait_s": 1.1733995052054524e-05, "gpu_peak_gb": 89.37899398803711}

runs/void-byte-v1/events.out.tfevents.1791463982.2e076f0c-3fe2-4316-8f17-4e30b2e706e2-b5972.360.0.1791463982175382982-d4a3b372

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b22a16d7058bb76031998fa5e186eeb13a5d304876d7e8a79d55ced9695a6b3a
-size 855747
+oid sha256:7cdd22f8d848699f8904d3c86bef0ec1e894ee5af032d5d8dbec1a149e8a11a8
+size 857865