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": 77557.50309181873,
-      "required_pages_per_second_upper_estimate": 0.005365608022092594
+      "required_token_positions_per_second": 76758.83112125874,
+      "required_pages_per_second_upper_estimate": 0.005310354041994965
     },
     {
       "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": 9694.68788647734,
-      "required_pages_per_second_upper_estimate": 0.004058963369963852
+      "required_token_positions_per_second": 9594.853890157343,
+      "required_pages_per_second_upper_estimate": 0.0040171649604755786
     },
     {
       "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": 19389.37577295468,
-      "required_pages_per_second_upper_estimate": 0.004038295641281725
+      "required_token_positions_per_second": 19189.707780314686,
+      "required_pages_per_second_upper_estimate": 0.0039967100640138755
     },
     {
       "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": 29084.063659432024,
-      "required_pages_per_second_upper_estimate": 0.0017435452075890946
+      "required_token_positions_per_second": 28784.56167047203,
+      "required_pages_per_second_upper_estimate": 0.0017255905206640008
     },
     {
       "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": 29084.063659432024,
-      "required_pages_per_second_upper_estimate": 0.0022611580222445735
+      "required_token_positions_per_second": 28784.56167047203,
+      "required_pages_per_second_upper_estimate": 0.0022378730599729594
     },
     {
       "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": 29084.063659432024,
-      "required_pages_per_second_upper_estimate": 0.3730591405886536
+      "required_token_positions_per_second": 28784.56167047203,
+      "required_pages_per_second_upper_estimate": 0.3692174506544558
     }
   ],
   "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": 2.558590814491007,
   "selected": {
-    "prefetch_batches": 64,
+    "prefetch_batches": 63,
     "prefetch_capacity_limit": 512,
     "batch_memory_estimate_bytes": 720750,
-    "producer_batch_seconds": 0.0858099744455103,
-    "producer_p95_seconds": 0.2837734869972337,
-    "gpu_update_seconds": 0.9506236928839101,
-    "producer_batches_per_second": 11.6536568908432,
-    "gpu_batches_per_second": 1.0519409599042253,
-    "producer_headroom": 11.078242349175389,
+    "producer_batch_seconds": 0.13635974532814998,
+    "producer_p95_seconds": 0.3058181180022075,
+    "gpu_update_seconds": 0.9605148869910378,
+    "producer_batches_per_second": 7.333542590546045,
+    "gpu_batches_per_second": 1.0411082780118646,
+    "producer_headroom": 7.043976832602297,
     "sustainable": true,
-    "buffer_seconds": 60.83991634457025,
+    "buffer_seconds": 60.512437880435385,
     "buffer_drain_seconds": null,
     "shard_target_mib": 16,
-    "upload_pages_per_commit": 84,
+    "upload_pages_per_commit": 83,
     "max_pending_pages": 4096,
-    "generation_bytes_per_second": 461898.06071651174,
+    "generation_bytes_per_second": 462930.00744985684,
     "upload_bytes_per_second": 4782377.5626485,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 99.11257139321818,
+    "estimated_full_shards_per_hour": 99.33400313970355,
     "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

@@ -342,3 +342,4 @@
 {"step": 919200, "loss": 0.5040311813354492, "lr": 3e-05, "grad_norm": 0.33464715896394387, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 134544169563, "tokens_per_second": 151711.25975651506, "data_wait_s": 1.4194003597367555e-05, "gpu_peak_gb": 89.37899398803711}
 {"step": 919300, "loss": 0.5025534629821777, "lr": 3e-05, "grad_norm": 0.3461679753224539, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 134558711089, "tokens_per_second": 152180.92482593792, "data_wait_s": 1.302100281463936e-05, "gpu_peak_gb": 89.37899398803711}
 {"step": 919400, "loss": 0.5069428086280823, "lr": 3e-05, "grad_norm": 0.3316022548419317, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 134573242699, "tokens_per_second": 151538.09558525475, "data_wait_s": 1.6842997865751386e-05, "gpu_peak_gb": 89.37899398803711}
+{"step": 919500, "loss": 0.549432098865509, "lr": 3e-05, "grad_norm": 0.3622988405517595, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 134587754931, "tokens_per_second": 151684.4907799723, "data_wait_s": 1.6425998182967305e-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 sha256:1de604391466cd32b70f16bfa9d02484831f18feee859503b57c2a4fafc984d6
-size 733455
+oid sha256:ab6e7e8cad93d392e4208400ce14068c82295f0c5a9565bc27243fbe74a4b8ca
+size 735573