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": 77373.54621660494,
-      "required_pages_per_second_upper_estimate": 0.005352881458626547
+      "required_token_positions_per_second": 77158.04572674513,
+      "required_pages_per_second_upper_estimate": 0.005337972634708018
     },
     {
       "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": 9671.693277075618,
-      "required_pages_per_second_upper_estimate": 0.004049336007189397
+      "required_token_positions_per_second": 9644.755715843141,
+      "required_pages_per_second_upper_estimate": 0.004038057812821605
     },
     {
       "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": 19343.386554151235,
-      "required_pages_per_second_upper_estimate": 0.004028717299822225
+      "required_token_positions_per_second": 19289.511431686282,
+      "required_pages_per_second_upper_estimate": 0.004017496532595298
     },
     {
       "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": 29015.079831226853,
-      "required_pages_per_second_upper_estimate": 0.0017394097324204012
+      "required_token_positions_per_second": 28934.26714752942,
+      "required_pages_per_second_upper_estimate": 0.0017345651354265634
     },
     {
       "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": 29015.079831226853,
-      "required_pages_per_second_upper_estimate": 0.0022557948330294143
+      "required_token_positions_per_second": 28934.26714752942,
+      "required_pages_per_second_upper_estimate": 0.0022495120023293684
     },
     {
       "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": 29015.079831226853,
-      "required_pages_per_second_upper_estimate": 0.3721742901095016
+      "required_token_positions_per_second": 28934.26714752942,
+      "required_pages_per_second_upper_estimate": 0.3711377117729303
     }
   ],
   "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": 0.11852353861729625,
-    "producer_p95_seconds": 0.3401618310017511,
-    "gpu_update_seconds": 0.9528838163059071,
-    "producer_batches_per_second": 8.437142627245766,
-    "gpu_batches_per_second": 1.0494458850993509,
-    "producer_headroom": 8.039616665367193,
+    "producer_batch_seconds": 0.12227666098408463,
+    "producer_p95_seconds": 0.35277252899686573,
+    "gpu_update_seconds": 0.9555451969468924,
+    "producer_batches_per_second": 8.178175556577871,
+    "gpu_batches_per_second": 1.0465229726392296,
+    "producer_headroom": 7.814616372876464,
     "sustainable": true,
-    "buffer_seconds": 60.03168042727215,
+    "buffer_seconds": 60.19934740765422,
     "buffer_drain_seconds": null,
     "shard_target_mib": 16,
-    "upload_pages_per_commit": 85,
+    "upload_pages_per_commit": 83,
     "max_pending_pages": 4096,
-    "generation_bytes_per_second": 320579.87488752574,
+    "generation_bytes_per_second": 297309.2619612056,
     "upload_bytes_per_second": 5106636.827372692,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 68.78897843331652,
+    "estimated_full_shards_per_hour": 63.79564661147237,
     "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

@@ -424,3 +424,4 @@
 {"step": 439500, "loss": 0.5806330442428589, "lr": 0.0003, "grad_norm": 0.16768765683487677, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 64330335108, "tokens_per_second": 153885.36132112876, "data_wait_s": 8.455004717689008e-06, "gpu_peak_gb": 89.37861251831055}
 {"step": 439600, "loss": 0.6042528748512268, "lr": 0.0003, "grad_norm": 0.13606011704472593, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 64344771197, "tokens_per_second": 150769.85554022, "data_wait_s": 1.493600575486198e-05, "gpu_peak_gb": 89.37861251831055}
 {"step": 439700, "loss": 0.5674126148223877, "lr": 0.0003, "grad_norm": 0.16720915015036775, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 64359282881, "tokens_per_second": 149590.2951846445, "data_wait_s": 1.2101998436264694e-05, "gpu_peak_gb": 89.37861251831055}
+{"step": 439800, "loss": 0.6283659338951111, "lr": 0.0003, "grad_norm": 0.1694328034722055, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 64373739148, "tokens_per_second": 149804.70826847616, "data_wait_s": 9.584997314959764e-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 sha256:3aa1e13ada9c03c2e90dcca2f805e6e329056ac99e0468b29e026d79eb8f374a
-size 908697
+oid sha256:7a3c9700567f7d22145df729074d8a8d3678f6b51b8b6f636c2d3340d9482818
+size 910815