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": 76033.02640911474,
-      "required_pages_per_second_upper_estimate": 0.006590350025198417
+      "required_token_positions_per_second": 76375.43340313055,
+      "required_pages_per_second_upper_estimate": 0.006620028995617115
     },
     {
       "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": 9504.128301139343,
-      "required_pages_per_second_upper_estimate": 0.0037913388787056577
+      "required_token_positions_per_second": 9546.929175391318,
+      "required_pages_per_second_upper_estimate": 0.0038084127873748677
     },
     {
       "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": 19008.256602278685,
-      "required_pages_per_second_upper_estimate": 0.004397225631046725
+      "required_token_positions_per_second": 19093.858350782637,
+      "required_pages_per_second_upper_estimate": 0.004417028089023795
     },
     {
       "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": 28512.384903418028,
-      "required_pages_per_second_upper_estimate": 0.0017362546476957198
+      "required_token_positions_per_second": 28640.787526173957,
+      "required_pages_per_second_upper_estimate": 0.0017440736937450589
     },
     {
       "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": 28512.384903418028,
-      "required_pages_per_second_upper_estimate": 0.0021949623171981067
+      "required_token_positions_per_second": 28640.787526173957,
+      "required_pages_per_second_upper_estimate": 0.002204847106539067
     },
     {
       "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": 22,
-      "required_token_positions_per_second": 28512.384903418028,
-      "required_pages_per_second_upper_estimate": 0.3540283956866785
+      "required_token_positions_per_second": 28640.787526173957,
+      "required_pages_per_second_upper_estimate": 0.3556227265253729
     }
   ],
   "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": 62,
+    "prefetch_batches": 63,
     "prefetch_capacity_limit": 512,
     "batch_memory_estimate_bytes": 688946,
-    "producer_batch_seconds": 0.08097002785922314,
-    "producer_p95_seconds": 0.2976478729979135,
-    "gpu_update_seconds": 0.9696838792564698,
-    "producer_batches_per_second": 12.350248930859074,
-    "gpu_batches_per_second": 1.031263921564599,
-    "producer_headroom": 11.975837293058495,
+    "producer_batch_seconds": 0.5055529695468408,
+    "producer_p95_seconds": 0.33400089199858485,
+    "gpu_update_seconds": 0.9653365842239262,
+    "producer_batches_per_second": 1.9780320960162956,
+    "gpu_batches_per_second": 1.0359081136492316,
+    "producer_headroom": 1.9094667470536641,
     "sustainable": true,
-    "buffer_seconds": 60.12040051390113,
+    "buffer_seconds": 60.81620480610735,
     "buffer_drain_seconds": null,
     "shard_target_mib": 16,
-    "upload_pages_per_commit": 82,
+    "upload_pages_per_commit": 80,
     "max_pending_pages": 4096,
-    "generation_bytes_per_second": 317636.3542804901,
+    "generation_bytes_per_second": 283448.8065840012,
     "upload_bytes_per_second": 4514531.554258477,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 68.15736743269946,
+    "estimated_full_shards_per_hour": 60.821515542412065,
     "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-v1/1790563279934888633-82973089/train.jsonl

@@ -297,3 +297,4 @@
 {"step": 184700, "loss": 0.5921379346400499, "lr": 0.0003, "grad_norm": 0.1538417757211502, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 27032382142, "tokens_per_second": 148764.65570608995, "data_wait_s": 1.221599814016372e-05, "gpu_peak_gb": 85.46427917480469}
 {"step": 184800, "loss": 0.6112805008888245, "lr": 0.0003, "grad_norm": 0.12806757929661203, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 27046955426, "tokens_per_second": 149143.50961878238, "data_wait_s": 8.838000212563202e-06, "gpu_peak_gb": 85.46427917480469}
 {"step": 184900, "loss": 0.5951752122491598, "lr": 0.0003, "grad_norm": 0.138714350567583, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 27061642898, "tokens_per_second": 129181.3069114033, "data_wait_s": 8.664999768370762e-06, "gpu_peak_gb": 85.46427917480469}
+{"step": 185000, "loss": 0.5817938577383757, "lr": 0.0003, "grad_norm": 0.151997777867797, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 27076235668, "tokens_per_second": 149048.55473883892, "data_wait_s": 9.453000529902056e-06, "gpu_peak_gb": 85.46427917480469}

runs/void-byte-v1/events.out.tfevents.1790563279.bdfa290c-aca9-4e96-a5d2-8425921a3278-d85hq.1304.0.1790563279934888633-82973089

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e57b8fd208457b8167d26e9c2767ef216c1b0cd49afde754032159f0b830f283
-size 637362
+oid sha256:68f96ff5f7c022af899099f41cfb21bc7c693d1f8ae268684e91e9b694ef7151
+size 639480