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": 74934.60173243853,
-      "required_pages_per_second_upper_estimate": 0.006495141358156014
+      "required_token_positions_per_second": 73204.54359674091,
+      "required_pages_per_second_upper_estimate": 0.006345184303746002
     },
     {
       "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": 9366.825216554817,
-      "required_pages_per_second_upper_estimate": 0.0037365666254008363
+      "required_token_positions_per_second": 9150.567949592614,
+      "required_pages_per_second_upper_estimate": 0.0036502983682753366
     },
     {
       "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": 18733.650433109633,
-      "required_pages_per_second_upper_estimate": 0.004333700326712978
+      "required_token_positions_per_second": 18301.13589918523,
+      "required_pages_per_second_upper_estimate": 0.004233645701285382
     },
     {
       "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": 28100.475649664448,
-      "required_pages_per_second_upper_estimate": 0.0017111715352629612
+      "required_token_positions_per_second": 27451.703848777845,
+      "required_pages_per_second_upper_estimate": 0.0016716647364315434
     },
     {
       "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": 28100.475649664448,
-      "required_pages_per_second_upper_estimate": 0.0021632524026063632
+      "required_token_positions_per_second": 27451.703848777845,
+      "required_pages_per_second_upper_estimate": 0.0021133081534588237
     },
     {
       "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": 21,
-      "required_token_positions_per_second": 28100.475649664448,
-      "required_pages_per_second_upper_estimate": 0.3489138613266505
+      "required_token_positions_per_second": 27451.703848777845,
+      "required_pages_per_second_upper_estimate": 0.3408582868591808
     }
   ],
   "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": 61,
+    "prefetch_batches": 60,
     "prefetch_capacity_limit": 512,
     "batch_memory_estimate_bytes": 688946,
-    "producer_batch_seconds": 0.09575027106257039,
-    "producer_p95_seconds": 0.36661074600124266,
-    "gpu_update_seconds": 0.9838979362731943,
-    "producer_batches_per_second": 10.443834663888577,
-    "gpu_batches_per_second": 1.0163655833935348,
-    "producer_headroom": 10.275667372578422,
+    "producer_batch_seconds": 0.11229810356286407,
+    "producer_p95_seconds": 0.356262464003521,
+    "gpu_update_seconds": 1.0071505999155275,
+    "producer_batches_per_second": 8.904869880017197,
+    "gpu_batches_per_second": 0.9929001681415598,
+    "producer_headroom": 8.96854504182903,
     "sustainable": true,
-    "buffer_seconds": 60.01777411266485,
+    "buffer_seconds": 60.429035994931645,
     "buffer_drain_seconds": null,
     "shard_target_mib": 16,
-    "upload_pages_per_commit": 75,
+    "upload_pages_per_commit": 84,
     "max_pending_pages": 4096,
-    "generation_bytes_per_second": 383003.6361229521,
+    "generation_bytes_per_second": 382009.22899517324,
     "upload_bytes_per_second": 5598717.511128207,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 82.1836644436495,
+    "estimated_full_shards_per_hour": 81.97028782264135,
     "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

@@ -359,3 +359,4 @@
 {"step": 75900, "loss": 0.5978897251188755, "lr": 0.0003, "grad_norm": 0.13849490037262083, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 11107875624, "tokens_per_second": 148662.9505346653, "data_wait_s": 1.391399564454332e-05, "gpu_peak_gb": 85.46427917480469}
 {"step": 76000, "loss": 0.6361755728721619, "lr": 0.0003, "grad_norm": 0.1380576848242837, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 11122456382, "tokens_per_second": 149707.9525421129, "data_wait_s": 1.0894000297412276e-05, "gpu_peak_gb": 85.46427917480469}
 {"step": 76100, "loss": 0.6172399986535311, "lr": 0.0003, "grad_norm": 0.13460179056605054, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 11137043847, "tokens_per_second": 149060.9724154184, "data_wait_s": 7.967995770741254e-06, "gpu_peak_gb": 85.46427917480469}
+{"step": 76200, "loss": 0.6052996497601271, "lr": 0.0003, "grad_norm": 0.14071265434358726, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 11151709472, "tokens_per_second": 150523.14569166646, "data_wait_s": 1.5386001905426383e-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 sha256:0cfb45712482cb579814cab8de2099596c9364af67de561cb50fd9dfa491916c
-size 770244
+oid sha256:ca4c64e0b4646b4c47e8a6397c91fc17f0bfa3b467bd5ebeff1f17f960d9d863
+size 772362