Update TensorBoard training metrics

benchmark_cache/shard_scheduler.json

@@ -22,8 +22,8 @@
       "tokens_per_second": 1062967.6168812835,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 1,
-      "required_token_positions_per_second": 73757.42736056469,
-      "required_pages_per_second_upper_estimate": 0.006393106867123158
+      "required_token_positions_per_second": 72607.03790201087,
+      "required_pages_per_second_upper_estimate": 0.00629339402449114
     },
     {
       "source": "rewrite6",
@@ -36,8 +36,8 @@
       "tokens_per_second": 1206672.263377472,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 1,
-      "required_token_positions_per_second": 9219.678420070586,
-      "required_pages_per_second_upper_estimate": 0.0036778675682426144
+      "required_token_positions_per_second": 9075.879737751358,
+      "required_pages_per_second_upper_estimate": 0.00362050412388358
     },
     {
       "source": "openmath",
@@ -50,8 +50,8 @@
       "tokens_per_second": 410042.7581134458,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 1,
-      "required_token_positions_per_second": 18439.356840141172,
-      "required_pages_per_second_upper_estimate": 0.004265620683370062
+      "required_token_positions_per_second": 18151.759475502717,
+      "required_pages_per_second_upper_estimate": 0.0041990900946017005
     },
     {
       "source": "ultra_qa",
@@ -64,8 +64,8 @@
       "tokens_per_second": 2163486.5624068812,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 1,
-      "required_token_positions_per_second": 27659.03526021176,
-      "required_pages_per_second_upper_estimate": 0.0016842901316040215
+      "required_token_positions_per_second": 27227.639213254075,
+      "required_pages_per_second_upper_estimate": 0.001658020375704435
     },
     {
       "source": "ultra_style",
@@ -78,8 +78,8 @@
       "tokens_per_second": 1546694.2737633225,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 1,
-      "required_token_positions_per_second": 27659.03526021176,
-      "required_pages_per_second_upper_estimate": 0.0021292690994410865
+      "required_token_positions_per_second": 27227.639213254075,
+      "required_pages_per_second_upper_estimate": 0.002096059037565595
     },
     {
       "source": "cortex",
@@ -92,8 +92,8 @@
       "tokens_per_second": 1015177.8395898147,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 21,
-      "required_token_positions_per_second": 27659.03526021176,
-      "required_pages_per_second_upper_estimate": 0.34343264909559285
+      "required_token_positions_per_second": 27227.639213254075,
+      "required_pages_per_second_upper_estimate": 0.3380761539820713
     }
   ],
   "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": 0.1434574083060457,
   "selected": {
-    "prefetch_batches": 61,
+    "prefetch_batches": 60,
     "prefetch_capacity_limit": 512,
     "batch_memory_estimate_bytes": 690670,
-    "producer_batch_seconds": 0.1862012017422785,
-    "producer_p95_seconds": 0.43837926800188143,
-    "gpu_update_seconds": 0.9996010251222451,
-    "producer_batches_per_second": 5.370534618697587,
-    "gpu_batches_per_second": 1.0003991341222425,
-    "producer_headroom": 5.368391910304613,
+    "producer_batch_seconds": 0.36094758826556017,
+    "producer_p95_seconds": 0.4999005789941293,
+    "gpu_update_seconds": 1.0154387526385797,
+    "producer_batches_per_second": 2.770485335018417,
+    "gpu_batches_per_second": 0.9847959784886458,
+    "producer_headroom": 2.813258172794579,
     "sustainable": true,
-    "buffer_seconds": 60.97566253245695,
+    "buffer_seconds": 60.92632515831478,
     "buffer_drain_seconds": null,
     "shard_target_mib": 16,
-    "upload_pages_per_commit": 83,
+    "upload_pages_per_commit": 78,
     "max_pending_pages": 4096,
-    "generation_bytes_per_second": 321576.6813675006,
+    "generation_bytes_per_second": 314137.48796374485,
     "upload_bytes_per_second": 5450583.341950757,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 69.00286989945187,
+    "estimated_full_shards_per_hour": 67.40659217056522,
     "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-v1/1790651964532391049-1e9227c4/train.jsonl

@@ -404,3 +404,4 @@
 {"step": 277500, "loss": 0.6091627143323421, "lr": 0.0003, "grad_norm": 0.15208406848093192, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 40617975581, "tokens_per_second": 149867.00935391747, "data_wait_s": 1.4083998394198716e-05, "gpu_peak_gb": 85.46427917480469}
 {"step": 277600, "loss": 0.598962739109993, "lr": 0.0003, "grad_norm": 0.16391701987067572, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 40632621246, "tokens_per_second": 148583.28325976295, "data_wait_s": 2.0689993107225746e-05, "gpu_peak_gb": 85.46427917480469}
 {"step": 277700, "loss": 0.5741484090685844, "lr": 0.0003, "grad_norm": 0.13775773153410134, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 40647240294, "tokens_per_second": 148081.28318683952, "data_wait_s": 1.3342003512661904e-05, "gpu_peak_gb": 85.46427917480469}
+{"step": 277800, "loss": 0.6059135496616364, "lr": 0.0003, "grad_norm": 0.1369716258194703, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 40662069546, "tokens_per_second": 147116.91008234757, "data_wait_s": 1.906400575535372e-05, "gpu_peak_gb": 85.46427917480469}

runs/void-byte-v1/events.out.tfevents.1790651964.96f3217e-af24-484e-a73b-abea9eb37bd2-c4mbg.442.0.1790651964532391049-1e9227c4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d8090d8e41e17f51c6f39e1cd0120904db45c93640e64754ce70e98c138d3d40
-size 866337
+oid sha256:6d3b44b4b2364ba0217ef76ebc3a1f47322c4e63be2f4f1986e9c829c3f30600
+size 868455