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": 77142.69758446155,
-      "required_pages_per_second_upper_estimate": 0.005336910814612243
+      "required_token_positions_per_second": 77299.33157173482,
+      "required_pages_per_second_upper_estimate": 0.005347747117292731
     },
     {
       "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": 9642.837198057694,
-      "required_pages_per_second_upper_estimate": 0.004037254569488045
+      "required_token_positions_per_second": 9662.416446466852,
+      "required_pages_per_second_upper_estimate": 0.0040454519919357595
     },
     {
       "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": 19285.674396115388,
-      "required_pages_per_second_upper_estimate": 0.004016697379275314
+      "required_token_positions_per_second": 19324.832892933704,
+      "required_pages_per_second_upper_estimate": 0.004024853061483563
     },
     {
       "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": 28928.51159417308,
-      "required_pages_per_second_upper_estimate": 0.001734220098791072
+      "required_token_positions_per_second": 28987.249339400554,
+      "required_pages_per_second_upper_estimate": 0.001737741336930115
     },
     {
       "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": 28928.51159417308,
-      "required_pages_per_second_upper_estimate": 0.002249064533371917
+      "required_token_positions_per_second": 28987.249339400554,
+      "required_pages_per_second_upper_estimate": 0.002253631134703316
     },
     {
       "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": 28928.51159417308,
-      "required_pages_per_second_upper_estimate": 0.3710638857143069
+      "required_token_positions_per_second": 28987.249339400554,
+      "required_pages_per_second_upper_estimate": 0.3718173104424078
     }
   ],
   "gpu_measurement": "Selected forward/backward estimate; replaced by live successful update timings.",
@@ -158,24 +158,24 @@
     "prefetch_batches": 63,
     "prefetch_capacity_limit": 512,
     "batch_memory_estimate_bytes": 720750,
-    "producer_batch_seconds": 0.10213259018757981,
-    "producer_p95_seconds": 0.22665445399979944,
-    "gpu_update_seconds": 0.9557353101280535,
-    "producer_batches_per_second": 9.791193958396333,
-    "gpu_batches_per_second": 1.0463148001364686,
-    "producer_headroom": 9.357789794351843,
+    "producer_batch_seconds": 0.06433898440629093,
+    "producer_p95_seconds": 0.2086363420003181,
+    "gpu_update_seconds": 0.9537986745924113,
+    "producer_batches_per_second": 15.542676174139642,
+    "gpu_batches_per_second": 1.048439284555865,
+    "producer_headroom": 14.82458393451344,
     "sustainable": true,
-    "buffer_seconds": 60.211324538067366,
+    "buffer_seconds": 60.08931649932191,
     "buffer_drain_seconds": null,
     "shard_target_mib": 16,
-    "upload_pages_per_commit": 83,
+    "upload_pages_per_commit": 85,
     "max_pending_pages": 4096,
-    "generation_bytes_per_second": 464530.8842065887,
+    "generation_bytes_per_second": 448035.51103309164,
     "upload_bytes_per_second": 5267230.115564385,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 99.6775140251946,
+    "estimated_full_shards_per_hour": 96.13799093479692,
     "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-v1/1791399463488568880-054ba646/train.jsonl

@@ -132,3 +132,4 @@
 {"step": 898200, "loss": 0.5086495876312256, "lr": 3e-05, "grad_norm": 0.3278603985220057, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 131470952327, "tokens_per_second": 151036.9326328292, "data_wait_s": 1.2135000361013226e-05, "gpu_peak_gb": 89.37899398803711}
 {"step": 898300, "loss": 0.559821367263794, "lr": 3e-05, "grad_norm": 0.3214628629964652, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 131485666654, "tokens_per_second": 153300.1605214211, "data_wait_s": 1.3654000213136896e-05, "gpu_peak_gb": 89.37899398803711}
 {"step": 898400, "loss": 0.5231099128723145, "lr": 3e-05, "grad_norm": 0.32097007019166945, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 131500370868, "tokens_per_second": 153466.88469262508, "data_wait_s": 1.2566999430418946e-05, "gpu_peak_gb": 89.37899398803711}
+{"step": 898500, "loss": 0.5456814765930176, "lr": 3e-05, "grad_norm": 0.35816702376225146, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 131515103268, "tokens_per_second": 153075.12642993315, "data_wait_s": 1.594099921931047e-05, "gpu_peak_gb": 89.37899398803711}

runs/void-byte-v1/events.out.tfevents.1791399463.60f38c85-66c9-4fe6-8e58-7b3a56dc2cee-wtkw8.346.0.1791399463488568880-054ba646

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:81bf358d700a84439a9c841c59b505f68ae9e5609817b3b0c2d4dd9c1b9dd22d
-size 285543
+oid sha256:5e66eb0a9389b3a45da3a37791843de7f9ae328f8bd9df1c9df1052a30b15a9f
+size 287661