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": 77523.71171552321,
-      "required_pages_per_second_upper_estimate": 0.005363270256273675
+      "required_token_positions_per_second": 77187.5334024352,
+      "required_pages_per_second_upper_estimate": 0.005340012660533092
     },
     {
       "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": 9690.463964440401,
-      "required_pages_per_second_upper_estimate": 0.004057194902012507
+      "required_token_positions_per_second": 9648.4416753044,
+      "required_pages_per_second_upper_estimate": 0.004039601047076448
     },
     {
       "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": 19380.927928880803,
-      "required_pages_per_second_upper_estimate": 0.004036536178145766
+      "required_token_positions_per_second": 19296.8833506088,
+      "required_pages_per_second_upper_estimate": 0.004019031908896283
     },
     {
       "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": 29071.391893321208,
-      "required_pages_per_second_upper_estimate": 0.0017427855545594177
+      "required_token_positions_per_second": 28945.325025913204,
+      "required_pages_per_second_upper_estimate": 0.0017352280383512685
     },
     {
       "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": 29071.391893321208,
-      "required_pages_per_second_upper_estimate": 0.002260172848166666
+      "required_token_positions_per_second": 28945.325025913204,
+      "required_pages_per_second_upper_estimate": 0.0022503717037352412
     },
     {
       "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": 29071.391893321208,
-      "required_pages_per_second_upper_estimate": 0.37289660077886644
+      "required_token_positions_per_second": 28945.325025913204,
+      "required_pages_per_second_upper_estimate": 0.3712795503638127
     }
   ],
   "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": 0.07703992790631276,
-    "producer_p95_seconds": 0.20368939800027874,
-    "gpu_update_seconds": 0.9510380549185808,
-    "producer_batches_per_second": 12.980282136505721,
-    "gpu_batches_per_second": 1.0514826350304256,
-    "producer_headroom": 12.3447422753968,
+    "producer_batch_seconds": 0.46438368425006615,
+    "producer_p95_seconds": 0.23935518100006448,
+    "gpu_update_seconds": 0.9551801534530437,
+    "producer_batches_per_second": 2.1533917618464598,
+    "gpu_batches_per_second": 1.0469229248377172,
+    "producer_headroom": 2.0568770735250217,
     "sustainable": true,
-    "buffer_seconds": 60.866435514789174,
+    "buffer_seconds": 60.17634966754176,
     "buffer_drain_seconds": null,
     "shard_target_mib": 16,
-    "upload_pages_per_commit": 88,
+    "upload_pages_per_commit": 78,
     "max_pending_pages": 4096,
-    "generation_bytes_per_second": 530569.8851616948,
+    "generation_bytes_per_second": 462284.5397538319,
     "upload_bytes_per_second": 5314690.275284022,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 113.84794632089742,
+    "estimated_full_shards_per_hour": 99.19550079785554,
     "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

@@ -118,3 +118,4 @@
 {"step": 896800, "loss": 0.5703634023666382, "lr": 3e-05, "grad_norm": 0.3258529454990818, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 131266109734, "tokens_per_second": 153753.41722248637, "data_wait_s": 1.816099938878324e-05, "gpu_peak_gb": 89.37899398803711}
 {"step": 896900, "loss": 0.5407653450965881, "lr": 3e-05, "grad_norm": 0.3022889709207004, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 131280877228, "tokens_per_second": 154489.044129573, "data_wait_s": 1.4236999049899168e-05, "gpu_peak_gb": 89.37899398803711}
 {"step": 897000, "loss": 0.5342614650726318, "lr": 3e-05, "grad_norm": 0.3248516682573976, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 131295500377, "tokens_per_second": 152090.90968864236, "data_wait_s": 1.5327999790315516e-05, "gpu_peak_gb": 89.37899398803711}
+{"step": 897100, "loss": 0.4904569387435913, "lr": 3e-05, "grad_norm": 0.33091313297640074, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 131310127884, "tokens_per_second": 153058.75319743843, "data_wait_s": 1.304299985349644e-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 sha256:20aafe80f34bbbcf7d11d4327a776d7a6f750837ae133f79e917e10ff168c24e
-size 255891
+oid sha256:eef3446246c6032a1f891a8140aff1d292fd8818b6435616dd2bc64f993935a7
+size 258009