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": 76178.60261053161,
-      "required_pages_per_second_upper_estimate": 0.005270212487307231
+      "required_token_positions_per_second": 74388.79837080979,
+      "required_pages_per_second_upper_estimate": 0.005146389677085279
     },
     {
       "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": 9522.325326316452,
-      "required_pages_per_second_upper_estimate": 0.003986798765364038
+      "required_token_positions_per_second": 9298.599796351224,
+      "required_pages_per_second_upper_estimate": 0.0038931295578879246
     },
     {
       "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": 19044.650652632903,
-      "required_pages_per_second_upper_estimate": 0.003966498489731465
+      "required_token_positions_per_second": 18597.199592702447,
+      "required_pages_per_second_upper_estimate": 0.0038733062340259223
     },
     {
       "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": 28566.975978949355,
-      "required_pages_per_second_upper_estimate": 0.0017125465906913302
+      "required_token_positions_per_second": 27895.79938905367,
+      "required_pages_per_second_upper_estimate": 0.0016723105789543974
     },
     {
       "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": 28566.975978949355,
-      "required_pages_per_second_upper_estimate": 0.002220956729515383
+      "required_token_positions_per_second": 27895.79938905367,
+      "required_pages_per_second_upper_estimate": 0.0021687757018448153
     },
     {
       "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": 22,
-      "required_token_positions_per_second": 28566.975978949355,
-      "required_pages_per_second_upper_estimate": 0.36642649502891644
+      "required_token_positions_per_second": 27895.79938905367,
+      "required_pages_per_second_upper_estimate": 0.3578173623870098
     }
   ],
   "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": 62,
+    "prefetch_batches": 61,
     "prefetch_capacity_limit": 512,
     "batch_memory_estimate_bytes": 720750,
-    "producer_batch_seconds": 0.06516900293755157,
-    "producer_p95_seconds": 0.22562466000090353,
-    "gpu_update_seconds": 0.967830827469224,
-    "producer_batches_per_second": 15.344718423239552,
-    "gpu_batches_per_second": 1.0332384251645454,
-    "producer_headroom": 14.851091528846181,
+    "producer_batch_seconds": 0.12734331103115437,
+    "producer_p95_seconds": 0.23342350500024622,
+    "gpu_update_seconds": 0.9911169640418727,
+    "producer_batches_per_second": 7.852787805677138,
+    "gpu_batches_per_second": 1.0089626515138046,
+    "producer_headroom": 7.783031209227764,
     "sustainable": true,
-    "buffer_seconds": 60.005511303091886,
+    "buffer_seconds": 60.458134806554234,
     "buffer_drain_seconds": null,
     "shard_target_mib": 16,
-    "upload_pages_per_commit": 85,
+    "upload_pages_per_commit": 86,
     "max_pending_pages": 4096,
-    "generation_bytes_per_second": 360657.48997459386,
+    "generation_bytes_per_second": 391420.0416455793,
     "upload_bytes_per_second": 3499768.2702418505,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 77.38870167187082,
+    "estimated_full_shards_per_hour": 83.98962914491209,
     "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-v1/1791342976705245012-6808dd50/train.jsonl

@@ -179,3 +179,4 @@
 {"step": 860400, "loss": 0.517084538936615, "lr": 3e-05, "grad_norm": 0.29384522981891675, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 125938188152, "tokens_per_second": 152395.935552823, "data_wait_s": 1.1955002264585346e-05, "gpu_peak_gb": 89.37899398803711}
 {"step": 860500, "loss": 0.5453771352767944, "lr": 3e-05, "grad_norm": 0.301904369976449, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 125952864797, "tokens_per_second": 152253.47331767945, "data_wait_s": 1.2029999197693542e-05, "gpu_peak_gb": 89.37899398803711}
 {"step": 860600, "loss": 0.5626469254493713, "lr": 3e-05, "grad_norm": 0.31876324271579914, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 125967547901, "tokens_per_second": 152391.45313855488, "data_wait_s": 1.5273002645699307e-05, "gpu_peak_gb": 89.37899398803711}
+{"step": 860700, "loss": 0.5463029742240906, "lr": 3e-05, "grad_norm": 0.30523915830765114, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 125981994594, "tokens_per_second": 143053.91110769127, "data_wait_s": 1.55200032168068e-05, "gpu_peak_gb": 89.37899398803711}

runs/void-byte-v1/events.out.tfevents.1791342976.f74d737f-395e-43a1-ae69-439b0359d002-2qp5c.418.0.1791342976705245012-6808dd50

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5a274c0c5df37650c0c50b52d86e1afdddf87667bb7c6ea3632ce576d543a895
-size 386655
+oid sha256:7b7f7371a98d79ed979b54384ecf9a9534c4c5e92313ab4719a67d501a515f23
+size 388773