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": 77173.57587575378,
-      "required_pages_per_second_upper_estimate": 0.0053390470464513475
+      "required_token_positions_per_second": 77482.48268197768,
+      "required_pages_per_second_upper_estimate": 0.005360417936068457
     },
     {
       "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": 9646.696984469223,
-      "required_pages_per_second_upper_estimate": 0.0040388705814570465
+      "required_token_positions_per_second": 9685.31033524721,
+      "required_pages_per_second_upper_estimate": 0.004055037185089334
     },
     {
       "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": 19293.393968938446,
-      "required_pages_per_second_upper_estimate": 0.004018305162715531
+      "required_token_positions_per_second": 19370.62067049442,
+      "required_pages_per_second_upper_estimate": 0.004034389448044564
     },
     {
       "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": 28940.09095340767,
-      "required_pages_per_second_upper_estimate": 0.0017349142636965279
+      "required_token_positions_per_second": 29055.931005741633,
+      "required_pages_per_second_upper_estimate": 0.0017418586979564305
     },
     {
       "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": 28940.09095340767,
-      "required_pages_per_second_upper_estimate": 0.0022499647776201876
+      "required_token_positions_per_second": 29055.931005741633,
+      "required_pages_per_second_upper_estimate": 0.0022589708321626117
     },
     {
       "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": 28940.09095340767,
-      "required_pages_per_second_upper_estimate": 0.37121241330162097
+      "required_token_positions_per_second": 29055.931005741633,
+      "required_pages_per_second_upper_estimate": 0.37269828511360337
     }
   ],
   "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": 63,
+    "prefetch_batches": 64,
     "prefetch_capacity_limit": 512,
     "batch_memory_estimate_bytes": 704566,
-    "producer_batch_seconds": 0.06801758105473965,
-    "producer_p95_seconds": 0.1473771510000006,
-    "gpu_update_seconds": 0.9553529062680597,
-    "producer_batches_per_second": 14.702081204493485,
-    "gpu_batches_per_second": 1.0467336137661918,
-    "producer_headroom": 14.045676006901866,
+    "producer_batch_seconds": 0.049011145750071705,
+    "producer_p95_seconds": 0.1542934340004649,
+    "gpu_update_seconds": 0.9515441096875052,
+    "producer_batches_per_second": 20.403522192674654,
+    "gpu_batches_per_second": 1.0509234304738726,
+    "producer_headroom": 19.414851359317858,
     "sustainable": true,
-    "buffer_seconds": 60.18723309488776,
+    "buffer_seconds": 60.898823020000336,
     "buffer_drain_seconds": null,
     "shard_target_mib": 16,
-    "upload_pages_per_commit": 79,
+    "upload_pages_per_commit": 82,
     "max_pending_pages": 4096,
-    "generation_bytes_per_second": 324227.4520590004,
+    "generation_bytes_per_second": 329488.8323514444,
     "upload_bytes_per_second": 5993507.971805119,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 69.57166358306416,
+    "estimated_full_shards_per_hour": 70.70063331515789,
     "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-v1/1790914434605211575-090606d4/train.jsonl

@@ -22,3 +22,4 @@
 {"step": 487200, "loss": 0.5762869119644165, "lr": 0.0003, "grad_norm": 0.1640187829808363, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 71310049986, "tokens_per_second": 152882.7269540695, "data_wait_s": 1.5609999536536634e-05, "gpu_peak_gb": 89.37899398803711}
 {"step": 487300, "loss": 0.6108181476593018, "lr": 0.0003, "grad_norm": 0.1651144827919894, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 71324568675, "tokens_per_second": 151824.51524321374, "data_wait_s": 7.669999831705354e-06, "gpu_peak_gb": 89.37899398803711}
 {"step": 487400, "loss": 0.5742596983909607, "lr": 0.0003, "grad_norm": 0.16172686769476063, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 71339094965, "tokens_per_second": 151759.73710918814, "data_wait_s": 9.094999768421985e-06, "gpu_peak_gb": 89.37899398803711}
+{"step": 487500, "loss": 0.5860530138015747, "lr": 0.0003, "grad_norm": 0.16962373406060718, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 71353551408, "tokens_per_second": 149781.97212477008, "data_wait_s": 1.2200000128359534e-05, "gpu_peak_gb": 89.37899398803711}

runs/void-byte-v1/events.out.tfevents.1790914434.4dec8a9d-3312-413a-9119-9058bec1adb7-bx5st.8214.0.1790914434605211575-090606d4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:bc8e18a629206d562421b24f84d60388fb2023922f577fb881498e5bfecb050e
-size 50997
+oid sha256:151f70321580d7b1c91f71a2d71af41f9f84874216b4a16b594f637b36e54089
+size 53115