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": 75393.27910870695,
-      "required_pages_per_second_upper_estimate": 0.006534898350623779
+      "required_token_positions_per_second": 76211.01170216338,
+      "required_pages_per_second_upper_estimate": 0.006605777339300272
     },
     {
       "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": 9424.159888588369,
-      "required_pages_per_second_upper_estimate": 0.0037594382833047585
+      "required_token_positions_per_second": 9526.376462770422,
+      "required_pages_per_second_upper_estimate": 0.0038002140030199546
     },
     {
       "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": 18848.319777176737,
-      "required_pages_per_second_upper_estimate": 0.004360227061351374
+      "required_token_positions_per_second": 19052.752925540844,
+      "required_pages_per_second_upper_estimate": 0.0044075190723248826
     },
     {
       "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": 28272.479665765106,
-      "required_pages_per_second_upper_estimate": 0.0017216456774081625
+      "required_token_positions_per_second": 28579.129388311267,
+      "required_pages_per_second_upper_estimate": 0.0017403190366444693
     },
     {
       "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": 28272.479665765106,
-      "required_pages_per_second_upper_estimate": 0.0021764937479033823
+      "required_token_positions_per_second": 28579.129388311267,
+      "required_pages_per_second_upper_estimate": 0.0022001004924057472
     },
     {
       "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": 28272.479665765106,
-      "required_pages_per_second_upper_estimate": 0.3510495755462099
+      "required_token_positions_per_second": 28579.129388311267,
+      "required_pages_per_second_upper_estimate": 0.35485713880963116
     }
   ],
   "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": 690670,
-    "producer_batch_seconds": 0.10294790071856141,
-    "producer_p95_seconds": 0.3309658430007403,
-    "gpu_update_seconds": 0.9779121013385578,
-    "producer_batches_per_second": 9.713651206291193,
-    "gpu_batches_per_second": 1.02258679346662,
-    "producer_headroom": 9.499097062814037,
+    "producer_batch_seconds": 0.12006928732853339,
+    "producer_p95_seconds": 0.3408958030049689,
+    "gpu_update_seconds": 0.9674192528519747,
+    "producer_batches_per_second": 8.328524489895585,
+    "gpu_batches_per_second": 1.0336780016026934,
+    "producer_headroom": 8.05717493937416,
     "sustainable": true,
-    "buffer_seconds": 60.63055028299058,
+    "buffer_seconds": 60.94741292967441,
     "buffer_drain_seconds": null,
     "shard_target_mib": 16,
-    "upload_pages_per_commit": 88,
+    "upload_pages_per_commit": 87,
     "max_pending_pages": 4096,
-    "generation_bytes_per_second": 315934.2535198036,
+    "generation_bytes_per_second": 316552.49004911387,
     "upload_bytes_per_second": 5614111.509302749,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 67.79213623233395,
+    "estimated_full_shards_per_hour": 67.92479539971411,
     "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-v1/1790703713325746110-adad6d18/train.jsonl

@@ -249,3 +249,4 @@
 {"step": 305100, "loss": 0.5901952311396599, "lr": 0.0003, "grad_norm": 0.1538424145560877, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 44658503019, "tokens_per_second": 126487.41882847535, "data_wait_s": 1.0969997674692422e-05, "gpu_peak_gb": 85.46991729736328}
 {"step": 305200, "loss": 0.6186392828822136, "lr": 0.0003, "grad_norm": 0.16531786503235366, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 44673192965, "tokens_per_second": 150494.23858165997, "data_wait_s": 7.946000550873578e-06, "gpu_peak_gb": 85.46991729736328}
 {"step": 305300, "loss": 0.592576090246439, "lr": 0.0003, "grad_norm": 0.16281926172888336, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 44687979341, "tokens_per_second": 151004.13191365328, "data_wait_s": 9.408999176230282e-06, "gpu_peak_gb": 85.46991729736328}
+{"step": 305400, "loss": 0.5878156870603561, "lr": 0.0003, "grad_norm": 0.14787342915316243, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 44702672078, "tokens_per_second": 151011.75474795562, "data_wait_s": 9.897994459606707e-06, "gpu_peak_gb": 85.46991729736328}

runs/void-byte-v1/events.out.tfevents.1790703713.3e6d470a-8cc7-4d7e-a224-51a0f9a2ce55-85btc.38629.0.1790703713325746110-adad6d18

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4fef0f0de2f1f438bddbcd28c09f5627bba985067972633b296d467ac5ffe6fa
-size 535698
+oid sha256:d2364d48a47c2b522fec50590087b8b683853f1c8f4221ecbc540c26bd537562
+size 537816