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": 77166.0012791878,
-      "required_pages_per_second_upper_estimate": 0.0053385230182854265
+      "required_token_positions_per_second": 76814.85012361473,
+      "required_pages_per_second_upper_estimate": 0.005314229566559946
     },
     {
       "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": 9645.750159898475,
-      "required_pages_per_second_upper_estimate": 0.00403847416578122
+      "required_token_positions_per_second": 9601.856265451841,
+      "required_pages_per_second_upper_estimate": 0.004020096708785161
     },
     {
       "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": 19291.50031979695,
-      "required_pages_per_second_upper_estimate": 0.0040179107655382434
+      "required_token_positions_per_second": 19203.712530903682,
+      "required_pages_per_second_upper_estimate": 0.003999626884231454
     },
     {
       "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": 28937.250479695424,
-      "required_pages_per_second_upper_estimate": 0.0017347439816346307
+      "required_token_positions_per_second": 28805.568796355525,
+      "required_pages_per_second_upper_estimate": 0.0017268498657846921
     },
     {
       "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": 28937.250479695424,
-      "required_pages_per_second_upper_estimate": 0.002249743943282984
+      "required_token_positions_per_second": 28805.568796355525,
+      "required_pages_per_second_upper_estimate": 0.0022395062716098193
     },
     {
       "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": 28937.250479695424,
-      "required_pages_per_second_upper_estimate": 0.3711759787547033
+      "required_token_positions_per_second": 28805.568796355525,
+      "required_pages_per_second_upper_estimate": 0.36948690750959484
     }
   ],
   "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": 0.1459351418360484,
-    "producer_p95_seconds": 0.3601418129983358,
-    "gpu_update_seconds": 0.9554466834850097,
-    "producer_batches_per_second": 6.852359119391924,
-    "gpu_batches_per_second": 1.046630876725095,
-    "producer_headroom": 6.547063794671275,
+    "producer_batch_seconds": 0.4687502106017405,
+    "producer_p95_seconds": 0.3604647050015046,
+    "gpu_update_seconds": 0.9598144093408084,
+    "producer_batches_per_second": 2.1333323748618427,
+    "gpu_batches_per_second": 1.0418680843589239,
+    "producer_headroom": 2.047603153305644,
     "sustainable": true,
-    "buffer_seconds": 60.193141059555614,
+    "buffer_seconds": 60.46830778847093,
     "buffer_drain_seconds": null,
     "shard_target_mib": 16,
-    "upload_pages_per_commit": 86,
+    "upload_pages_per_commit": 81,
     "max_pending_pages": 4096,
-    "generation_bytes_per_second": 358018.5087519352,
+    "generation_bytes_per_second": 324579.371270228,
     "upload_bytes_per_second": 5073450.113888784,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 76.82243773382704,
+    "estimated_full_shards_per_hour": 69.64717725353366,
     "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-v1/1790824801708454555-41033672/train.jsonl

@@ -403,3 +403,4 @@
 {"step": 437400, "loss": 0.6454831957817078, "lr": 0.0003, "grad_norm": 0.15734464689470082, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 64022942316, "tokens_per_second": 153265.52283383004, "data_wait_s": 1.1652002285700291e-05, "gpu_peak_gb": 89.37861251831055}
 {"step": 437500, "loss": 0.567116379737854, "lr": 0.0003, "grad_norm": 0.1992623283552305, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 64037726064, "tokens_per_second": 153628.012736391, "data_wait_s": 1.5614001313224435e-05, "gpu_peak_gb": 89.37861251831055}
 {"step": 437600, "loss": 0.5801525712013245, "lr": 0.0003, "grad_norm": 0.14818312405547457, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 64052297925, "tokens_per_second": 151189.28094464281, "data_wait_s": 1.2931006494909525e-05, "gpu_peak_gb": 89.37861251831055}
+{"step": 437700, "loss": 0.6120076179504395, "lr": 0.0003, "grad_norm": 0.1622745687217743, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 64066729368, "tokens_per_second": 137251.91893355988, "data_wait_s": 9.629999112803489e-06, "gpu_peak_gb": 89.37861251831055}

runs/void-byte-v1/events.out.tfevents.1790824801.afc92292-ba09-4e51-8d82-bc1983323308-sdcdh.469.0.1790824801708454555-41033672

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f9da875071f9ff2a949225ddc9b3902908452a3bf017cec4da35adbc08e6ff86
-size 864219
+oid sha256:b47df87eb3e1157f8f483255b93068271ba8bcf21b3823544a24b7b3cba55c5a
+size 866337