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": 76056.4175896576,
-      "required_pages_per_second_upper_estimate": 0.006592377513443467
+      "required_token_positions_per_second": 75523.3719819806,
+      "required_pages_per_second_upper_estimate": 0.006546174471161766
     },
     {
       "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": 9507.0521987072,
-      "required_pages_per_second_upper_estimate": 0.0037925052651616407
+      "required_token_positions_per_second": 9440.421497747575,
+      "required_pages_per_second_upper_estimate": 0.0037659252823310893
     },
     {
       "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": 19014.1043974144,
-      "required_pages_per_second_upper_estimate": 0.004398578415533694
+      "required_token_positions_per_second": 18880.84299549515,
+      "required_pages_per_second_upper_estimate": 0.0043677507355202455
     },
     {
       "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": 28521.156596121604,
-      "required_pages_per_second_upper_estimate": 0.0017367887977598527
+      "required_token_positions_per_second": 28321.264493242725,
+      "required_pages_per_second_upper_estimate": 0.001724616417447341
     },
     {
       "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": 28521.156596121604,
-      "required_pages_per_second_upper_estimate": 0.002195637586384027
+      "required_token_positions_per_second": 28321.264493242725,
+      "required_pages_per_second_upper_estimate": 0.002180249338967655
     },
     {
       "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": 28521.156596121604,
-      "required_pages_per_second_upper_estimate": 0.3541373107530899
+      "required_token_positions_per_second": 28321.264493242725,
+      "required_pages_per_second_upper_estimate": 0.3516553198311673
     }
   ],
   "gpu_measurement": "Selected forward/backward estimate; replaced by live successful update timings.",
@@ -158,24 +158,24 @@
     "prefetch_batches": 62,
     "prefetch_capacity_limit": 512,
     "batch_memory_estimate_bytes": 688946,
-    "producer_batch_seconds": 0.42189215079707765,
-    "producer_p95_seconds": 0.3517229930002941,
-    "gpu_update_seconds": 0.9693856526056752,
-    "producer_batches_per_second": 2.37027400986415,
-    "gpu_batches_per_second": 1.0315811847555558,
-    "producer_headroom": 2.29770961790643,
+    "producer_batch_seconds": 0.13910102510953948,
+    "producer_p95_seconds": 0.3497807180028758,
+    "gpu_update_seconds": 0.9762275976977172,
+    "producer_batches_per_second": 7.189019629528384,
+    "gpu_batches_per_second": 1.024351290988235,
+    "producer_headroom": 7.018119362736227,
     "sustainable": true,
-    "buffer_seconds": 60.10191046155187,
+    "buffer_seconds": 60.52611105725846,
     "buffer_drain_seconds": null,
     "shard_target_mib": 16,
-    "upload_pages_per_commit": 83,
+    "upload_pages_per_commit": 74,
     "max_pending_pages": 4096,
-    "generation_bytes_per_second": 316712.9985123715,
+    "generation_bytes_per_second": 303740.1106864649,
     "upload_bytes_per_second": 3255984.9248120873,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 67.9592367794834,
+    "estimated_full_shards_per_hour": 65.17555704541645,
     "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

@@ -289,3 +289,4 @@
 {"step": 68900, "loss": 0.6303941160440445, "lr": 0.0003, "grad_norm": 0.13800357157853751, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 10083778947, "tokens_per_second": 149816.38426153688, "data_wait_s": 1.1046999134123325e-05, "gpu_peak_gb": 85.46427917480469}
 {"step": 69000, "loss": 0.5993391498923302, "lr": 0.0003, "grad_norm": 0.11908661536376573, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 10098384185, "tokens_per_second": 150008.19648787344, "data_wait_s": 9.00599843589589e-06, "gpu_peak_gb": 85.46427917480469}
 {"step": 69100, "loss": 0.6415554154664278, "lr": 0.0003, "grad_norm": 0.12543848502533203, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 10113137289, "tokens_per_second": 149927.10782188523, "data_wait_s": 7.864997314754874e-06, "gpu_peak_gb": 85.46427917480469}
+{"step": 69200, "loss": 0.6179433986544609, "lr": 0.0003, "grad_norm": 0.13127947763161085, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 10127801907, "tokens_per_second": 150235.5464349099, "data_wait_s": 1.146899739978835e-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:a53fe1cb49412b878fb9622eb87a5795380a89a2ab02c1e3a9f7cf3b061a33bb
-size 620418
+oid sha256:450bc9e51a1476288122e99e210a4cc86db540ab84d66d7483f95e2c6ab10ec7
+size 622536