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": 75182.4714361772,
-      "required_pages_per_second_upper_estimate": 0.006516626075856062
+      "required_token_positions_per_second": 76337.14738157115,
+      "required_pages_per_second_upper_estimate": 0.006616710460303381
     },
     {
       "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": 9397.80892952215,
-      "required_pages_per_second_upper_estimate": 0.003748926491751297
+      "required_token_positions_per_second": 9542.143422696394,
+      "required_pages_per_second_upper_estimate": 0.003806503679071483
     },
     {
       "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": 18795.6178590443,
-      "required_pages_per_second_upper_estimate": 0.004348035400113515
+      "required_token_positions_per_second": 19084.286845392788,
+      "required_pages_per_second_upper_estimate": 0.004414813889704595
     },
     {
       "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": 28193.426788566452,
-      "required_pages_per_second_upper_estimate": 0.001716831771945157
+      "required_token_positions_per_second": 28626.430268089178,
+      "required_pages_per_second_upper_estimate": 0.0017431994120544075
     },
     {
       "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": 28193.426788566452,
-      "required_pages_per_second_upper_estimate": 0.0021704080385842603
+      "required_token_positions_per_second": 28626.430268089178,
+      "required_pages_per_second_upper_estimate": 0.0022037418450682666
     },
     {
       "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": 28193.426788566452,
-      "required_pages_per_second_upper_estimate": 0.3500680033843631
+      "required_token_positions_per_second": 28626.430268089178,
+      "required_pages_per_second_upper_estimate": 0.35544445743061176
     }
   ],
   "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": 62,
+    "prefetch_batches": 63,
     "prefetch_capacity_limit": 512,
     "batch_memory_estimate_bytes": 688946,
-    "producer_batch_seconds": 0.1010180596954342,
-    "producer_p95_seconds": 0.32333155699961935,
-    "gpu_update_seconds": 0.98065411513624,
-    "producer_batches_per_second": 9.899220030705042,
-    "gpu_batches_per_second": 1.019727531415164,
-    "producer_headroom": 9.707710859749996,
+    "producer_batch_seconds": 0.35829614098443585,
+    "producer_p95_seconds": 0.318228405998525,
+    "gpu_update_seconds": 0.9658207377264266,
+    "producer_batches_per_second": 2.7909873582574796,
+    "gpu_batches_per_second": 1.0353888262474384,
+    "producer_headroom": 2.695593469337369,
     "sustainable": true,
-    "buffer_seconds": 60.80055513844688,
+    "buffer_seconds": 60.84670647676488,
     "buffer_drain_seconds": null,
     "shard_target_mib": 16,
-    "upload_pages_per_commit": 88,
+    "upload_pages_per_commit": 84,
     "max_pending_pages": 4096,
-    "generation_bytes_per_second": 389209.6423848629,
+    "generation_bytes_per_second": 359803.62870726065,
     "upload_bytes_per_second": 3399051.8321701875,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 83.51532891902366,
+    "estimated_full_shards_per_hour": 77.20548292077412,
     "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-v1/1790434090210070261-a05fc3fb/train.jsonl

@@ -273,3 +273,4 @@
 {"step": 67300, "loss": 0.618693120777607, "lr": 0.0003, "grad_norm": 0.13847754745698787, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 9849572397, "tokens_per_second": 150969.64528632563, "data_wait_s": 7.422000635415316e-06, "gpu_peak_gb": 85.46427917480469}
 {"step": 67400, "loss": 0.6169149242341518, "lr": 0.0003, "grad_norm": 0.12542056418614891, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 9864109107, "tokens_per_second": 148125.61582019657, "data_wait_s": 9.758001397131011e-06, "gpu_peak_gb": 85.46427917480469}
 {"step": 67500, "loss": 0.6437095645815134, "lr": 0.0003, "grad_norm": 0.13284275550097244, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 9878791847, "tokens_per_second": 150492.14726144756, "data_wait_s": 7.589998858748004e-06, "gpu_peak_gb": 85.46427917480469}
+{"step": 67600, "loss": 0.6464938297867775, "lr": 0.0003, "grad_norm": 0.14398202691306705, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 9893419463, "tokens_per_second": 148779.30196316144, "data_wait_s": 1.2276999768801033e-05, "gpu_peak_gb": 85.46427917480469}

runs/void-byte-v1/events.out.tfevents.1790434090.372bd512-57cc-4928-b44a-3b375a0fa071-5spg5.451.0.1790434090210070261-a05fc3fb

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3880ef41e8771e90f4d86d529fce502e8070eccbbbacac2e317ca94ffa4e6f85
-size 586530
+oid sha256:e061059b5cad78a22e40a9b99d826ea844f7d4e4201e931a72d57c6aefad6fe0
+size 588648