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": 76415.45644689692,
-      "required_pages_per_second_upper_estimate": 0.0052865985852836
+      "required_token_positions_per_second": 76770.17518983975,
+      "required_pages_per_second_upper_estimate": 0.005311138850981265
     },
     {
       "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": 9551.932055862115,
-      "required_pages_per_second_upper_estimate": 0.00399919448476599
+      "required_token_positions_per_second": 9596.271898729969,
+      "required_pages_per_second_upper_estimate": 0.004017758651053551
     },
     {
       "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": 19103.86411172423,
-      "required_pages_per_second_upper_estimate": 0.003978831091696261
+      "required_token_positions_per_second": 19192.543797459937,
+      "required_pages_per_second_upper_estimate": 0.00399730073159443
     },
     {
       "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": 28655.796167586344,
-      "required_pages_per_second_upper_estimate": 0.0017178712253795424
+      "required_token_positions_per_second": 28788.815696189904,
+      "required_pages_per_second_upper_estimate": 0.001725845543010266
     },
     {
       "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": 28655.796167586344,
-      "required_pages_per_second_upper_estimate": 0.0022278621085031875
+      "required_token_positions_per_second": 28788.815696189904,
+      "required_pages_per_second_upper_estimate": 0.0022382037917610424
     },
     {
       "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": 28655.796167586344,
-      "required_pages_per_second_upper_estimate": 0.3675657850410634
+      "required_token_positions_per_second": 28788.815696189904,
+      "required_pages_per_second_upper_estimate": 0.36927201672874777
     }
   ],
   "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": 704566,
-    "producer_batch_seconds": 0.07967765762481349,
-    "producer_p95_seconds": 0.29691396099951817,
-    "gpu_update_seconds": 0.9648309835227575,
-    "producer_batches_per_second": 12.550569755812909,
-    "gpu_batches_per_second": 1.0364509609225385,
-    "producer_headroom": 12.109178561271943,
+    "producer_batch_seconds": 0.12358868788271593,
+    "producer_p95_seconds": 0.29516289200182655,
+    "gpu_update_seconds": 0.960372954961781,
+    "producer_batches_per_second": 8.091355423637049,
+    "gpu_batches_per_second": 1.041262141789276,
+    "producer_headroom": 7.770718917844346,
     "sustainable": true,
-    "buffer_seconds": 60.78435196193372,
+    "buffer_seconds": 60.5034961625922,
     "buffer_drain_seconds": null,
     "shard_target_mib": 16,
-    "upload_pages_per_commit": 85,
+    "upload_pages_per_commit": 80,
     "max_pending_pages": 4096,
-    "generation_bytes_per_second": 312820.428560372,
+    "generation_bytes_per_second": 327623.9002614106,
     "upload_bytes_per_second": 5818964.437699522,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 67.12398188217516,
+    "estimated_full_shards_per_hour": 70.3004623020338,
     "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-v1/1790989041640695387-36510d53/train.jsonl

@@ -285,3 +285,4 @@
 {"step": 585600, "loss": 0.5698740482330322, "lr": 0.0003, "grad_norm": 0.18035433676959006, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 85714511432, "tokens_per_second": 150697.80202270712, "data_wait_s": 1.2169002729933709e-05, "gpu_peak_gb": 89.37899398803711}
 {"step": 585700, "loss": 0.5852900743484497, "lr": 0.0003, "grad_norm": 0.1548418470033908, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 85729205185, "tokens_per_second": 145548.49376442036, "data_wait_s": 1.0175001079915091e-05, "gpu_peak_gb": 89.37899398803711}
 {"step": 585800, "loss": 0.5899698138237, "lr": 0.0003, "grad_norm": 0.18886438473564646, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 85743872795, "tokens_per_second": 152713.66823711075, "data_wait_s": 1.5837002138141543e-05, "gpu_peak_gb": 89.37899398803711}
+{"step": 585900, "loss": 0.5797560214996338, "lr": 0.0003, "grad_norm": 0.14082809131657953, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 85758494033, "tokens_per_second": 151858.86171897678, "data_wait_s": 6.595997547265142e-06, "gpu_peak_gb": 89.37899398803711}

runs/void-byte-v1/events.out.tfevents.1790989041.fb9d4967-10f3-4fa9-9622-1d5a28f8589b-vh8db.427.0.1790989041640695387-36510d53

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6d6d290015b12c338e661af0d710ce18ca25c7487c9933fcaf56f253856bd1d3
-size 612729
+oid sha256:0e64e38136d0f0616c678e078ef7057c9013f032f6989d33cb39ec5cf1920b82
+size 614847