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": 76820.44060621821,
-      "required_pages_per_second_upper_estimate": 0.005314616329118161
+      "required_token_positions_per_second": 76946.46270481875,
+      "required_pages_per_second_upper_estimate": 0.005323334830310902
     },
     {
       "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": 9602.555075777276,
-      "required_pages_per_second_upper_estimate": 0.004020389286075602
+      "required_token_positions_per_second": 9618.307838102344,
+      "required_pages_per_second_upper_estimate": 0.004026984638706024
     },
     {
       "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": 19205.110151554552,
-      "required_pages_per_second_upper_estimate": 0.0039999179717552956
+      "required_token_positions_per_second": 19236.61567620469,
+      "required_pages_per_second_upper_estimate": 0.004006479741683361
     },
     {
       "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": 28807.665227331832,
-      "required_pages_per_second_upper_estimate": 0.0017269755436206571
+      "required_token_positions_per_second": 28854.92351430703,
+      "required_pages_per_second_upper_estimate": 0.0017298086057655941
     },
     {
       "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": 28807.665227331832,
-      "required_pages_per_second_upper_estimate": 0.002239669259897002
+      "required_token_positions_per_second": 28854.92351430703,
+      "required_pages_per_second_upper_estimate": 0.002243343383842087
     },
     {
       "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": 28807.665227331832,
-      "required_pages_per_second_upper_estimate": 0.3695137982751867
+      "required_token_positions_per_second": 28854.92351430703,
+      "required_pages_per_second_upper_estimate": 0.3701199768385094
     }
   ],
   "gpu_measurement": "Selected forward/backward estimate; replaced by live successful update timings.",
@@ -158,24 +158,24 @@
     "prefetch_batches": 63,
     "prefetch_capacity_limit": 512,
     "batch_memory_estimate_bytes": 704566,
-    "producer_batch_seconds": 0.44656508080470303,
-    "producer_p95_seconds": 0.2887253939989023,
-    "gpu_update_seconds": 0.9597445604084716,
-    "producer_batches_per_second": 2.2393152599348256,
-    "gpu_batches_per_second": 1.041943910132083,
-    "producer_headroom": 2.1491706397621315,
+    "producer_batch_seconds": 0.27080134640624465,
+    "producer_p95_seconds": 0.2872216209980252,
+    "gpu_update_seconds": 0.9581727009704737,
+    "producer_batches_per_second": 3.692743825947758,
+    "gpu_batches_per_second": 1.0436531942385356,
+    "producer_headroom": 3.5382863257004042,
     "sustainable": true,
-    "buffer_seconds": 60.463907305733706,
+    "buffer_seconds": 60.364880161139844,
     "buffer_drain_seconds": null,
     "shard_target_mib": 16,
-    "upload_pages_per_commit": 82,
+    "upload_pages_per_commit": 78,
     "max_pending_pages": 4096,
-    "generation_bytes_per_second": 326334.7211430822,
+    "generation_bytes_per_second": 322056.5674674666,
     "upload_bytes_per_second": 4952883.252326766,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 70.02383447379445,
+    "estimated_full_shards_per_hour": 69.1058422853279,
     "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-v1/1790824801708454555-41033672/train.jsonl

@@ -238,3 +238,4 @@
 {"step": 420900, "loss": 0.5789002180099487, "lr": 0.0003, "grad_norm": 0.1462506781190776, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 61608463519, "tokens_per_second": 150293.09301309913, "data_wait_s": 1.3267999747768044e-05, "gpu_peak_gb": 89.37861251831055}
 {"step": 421000, "loss": 0.577822744846344, "lr": 0.0003, "grad_norm": 0.17032858961319572, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 61623016538, "tokens_per_second": 150942.25230837657, "data_wait_s": 1.3732998922932893e-05, "gpu_peak_gb": 89.37861251831055}
 {"step": 421100, "loss": 0.5536261796951294, "lr": 0.0003, "grad_norm": 0.1453920134542661, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 61637658452, "tokens_per_second": 151558.5535397857, "data_wait_s": 1.1459000234026462e-05, "gpu_peak_gb": 89.37861251831055}
+{"step": 421200, "loss": 0.5926738381385803, "lr": 0.0003, "grad_norm": 0.18534939670965112, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 61652404313, "tokens_per_second": 152882.80334982387, "data_wait_s": 9.011000656755641e-06, "gpu_peak_gb": 89.37861251831055}

runs/void-byte-v1/events.out.tfevents.1790824801.afc92292-ba09-4e51-8d82-bc1983323308-sdcdh.469.0.1790824801708454555-41033672

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:24a904efa72d9e43266bb797b8fda094d62adb284594a3f9132492978d7c6af3
-size 512400
+oid sha256:1ddcee427e1262f4f0f37f05a422cb8ca6c50be4555cad32948c2d7216668afe
+size 514518