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": 77501.08979938099,
-      "required_pages_per_second_upper_estimate": 0.0053617052196248795
+      "required_token_positions_per_second": 76810.08468874969,
+      "required_pages_per_second_upper_estimate": 0.00531389988271865
     },
     {
       "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": 9687.636224922624,
-      "required_pages_per_second_upper_estimate": 0.00405601098652633
+      "required_token_positions_per_second": 9601.260586093711,
+      "required_pages_per_second_upper_estimate": 0.004019847310277111
     },
     {
       "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": 19375.272449845248,
-      "required_pages_per_second_upper_estimate": 0.004035358291007671
+      "required_token_positions_per_second": 19202.521172187422,
+      "required_pages_per_second_upper_estimate": 0.003999378755629099
     },
     {
       "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": 29062.90867476787,
-      "required_pages_per_second_upper_estimate": 0.0017422769985602697
+      "required_token_positions_per_second": 28803.78175828113,
+      "required_pages_per_second_upper_estimate": 0.0017267427355807815
     },
     {
       "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": 29062.90867476787,
-      "required_pages_per_second_upper_estimate": 0.002259513315237879
+      "required_token_positions_per_second": 28803.78175828113,
+      "required_pages_per_second_upper_estimate": 0.002239367337259885
     },
     {
       "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": 29062.90867476787,
-      "required_pages_per_second_upper_estimate": 0.3727877871598347
+      "required_token_positions_per_second": 28803.78175828113,
+      "required_pages_per_second_upper_estimate": 0.36946398530394853
     }
   ],
   "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": 0.11459342196883426,
-    "producer_p95_seconds": 0.3316530660013086,
-    "gpu_update_seconds": 0.9513156549262985,
-    "producer_batches_per_second": 8.726504391080738,
-    "gpu_batches_per_second": 1.0511758056556666,
-    "producer_headroom": 8.301660240018192,
+    "producer_batch_seconds": 0.10954868127328155,
+    "producer_p95_seconds": 0.31392763199983165,
+    "gpu_update_seconds": 0.9598739579413442,
+    "producer_batches_per_second": 9.128361823958311,
+    "gpu_batches_per_second": 1.0418034490119044,
+    "producer_headroom": 8.762076793483532,
     "sustainable": true,
-    "buffer_seconds": 60.88420191528311,
+    "buffer_seconds": 60.47205935030468,
     "buffer_drain_seconds": null,
     "shard_target_mib": 16,
-    "upload_pages_per_commit": 86,
+    "upload_pages_per_commit": 81,
     "max_pending_pages": 4096,
-    "generation_bytes_per_second": 424633.1298277971,
+    "generation_bytes_per_second": 377443.02657550055,
     "upload_bytes_per_second": 5564620.844387995,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 91.11638470769344,
+    "estimated_full_shards_per_hour": 80.99048707913173,
     "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-v1/1791143623306220418-eb05a537/train.jsonl

@@ -377,3 +377,4 @@
 {"step": 722700, "loss": 0.5891305804252625, "lr": 8e-05, "grad_norm": 0.23494695261280726, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 105783284226, "tokens_per_second": 151962.07370628853, "data_wait_s": 9.780000254977494e-06, "gpu_peak_gb": 89.37899398803711}
 {"step": 722800, "loss": 0.5541865229606628, "lr": 8e-05, "grad_norm": 0.21566273233327088, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 105798063516, "tokens_per_second": 154729.68355813564, "data_wait_s": 1.4995996025390923e-05, "gpu_peak_gb": 89.37899398803711}
 {"step": 722900, "loss": 0.5568122863769531, "lr": 8e-05, "grad_norm": 0.21909042152464192, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 105812839171, "tokens_per_second": 153428.21704980786, "data_wait_s": 1.3438999303616583e-05, "gpu_peak_gb": 89.37899398803711}
+{"step": 723000, "loss": 0.5880641341209412, "lr": 8e-05, "grad_norm": 0.2674153980514474, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 105827552658, "tokens_per_second": 152964.67979861106, "data_wait_s": 1.0994001058861613e-05, "gpu_peak_gb": 89.37899398803711}

runs/void-byte-v1/events.out.tfevents.1791143623.3eb6ddb3-ffce-42e5-bd48-c0d653189e98-8pb7d.369.0.1791143623306220418-eb05a537

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:82e657aa5c6108b6c9aaccd9e978c95b0d51571eff0b1c546102d170c8b81713
-size 808368
+oid sha256:12931522cbd1582a2d6485f7498f367bbd9dc419bed37041d5e5a18ea096b781
+size 810486