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": 76226.80839679117,
-      "required_pages_per_second_upper_estimate": 0.006607146556754252
+      "required_token_positions_per_second": 75745.6130420059,
+      "required_pages_per_second_upper_estimate": 0.006565437762979943
     },
     {
       "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": 9528.351049598896,
-      "required_pages_per_second_upper_estimate": 0.0038010016952285367
+      "required_token_positions_per_second": 9468.201630250738,
+      "required_pages_per_second_upper_estimate": 0.003777007192536596
     },
     {
       "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": 19056.70209919779,
-      "required_pages_per_second_upper_estimate": 0.004408432644147335
+      "required_token_positions_per_second": 18936.403260501476,
+      "required_pages_per_second_upper_estimate": 0.004380603625002192
     },
     {
       "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": 28585.053148796687,
-      "required_pages_per_second_upper_estimate": 0.0017406797625784621
+      "required_token_positions_per_second": 28404.604890752213,
+      "required_pages_per_second_upper_estimate": 0.0017296914104023957
     },
     {
       "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": 28585.053148796687,
-      "required_pages_per_second_upper_estimate": 0.0022005565198859336
+      "required_token_positions_per_second": 28404.604890752213,
+      "required_pages_per_second_upper_estimate": 0.0021866651134689213
     },
     {
       "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": 28585.053148796687,
-      "required_pages_per_second_upper_estimate": 0.3549306920893091
+      "required_token_positions_per_second": 28404.604890752213,
+      "required_pages_per_second_upper_estimate": 0.3526901286458673
     }
   ],
   "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": 63,
+    "prefetch_batches": 62,
     "prefetch_capacity_limit": 512,
     "batch_memory_estimate_bytes": 690670,
-    "producer_batch_seconds": 0.32516137095308295,
-    "producer_p95_seconds": 0.34261003600113327,
-    "gpu_update_seconds": 0.9672187718553837,
-    "producer_batches_per_second": 3.075396062788432,
-    "gpu_batches_per_second": 1.0338922579859913,
-    "producer_headroom": 2.9745808028191094,
+    "producer_batch_seconds": 0.4270988903281534,
+    "producer_p95_seconds": 0.34832001500035403,
+    "gpu_update_seconds": 0.973363301701882,
+    "producer_batches_per_second": 2.341378127280239,
+    "gpu_batches_per_second": 1.027365628282415,
+    "producer_headroom": 2.2790115445020627,
     "sustainable": true,
-    "buffer_seconds": 60.934782626889174,
+    "buffer_seconds": 60.34852470551669,
     "buffer_drain_seconds": null,
     "shard_target_mib": 16,
-    "upload_pages_per_commit": 83,
+    "upload_pages_per_commit": 79,
     "max_pending_pages": 4096,
-    "generation_bytes_per_second": 399593.1058812843,
+    "generation_bytes_per_second": 364468.8788837926,
     "upload_bytes_per_second": 3319725.431627655,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 85.74337847069641,
+    "estimated_full_shards_per_hour": 78.20653700719198,
     "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-v1/1790703713325746110-adad6d18/train.jsonl

@@ -230,3 +230,4 @@
 {"step": 303200, "loss": 0.7135458812117577, "lr": 0.0003, "grad_norm": 0.22248423776876727, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 44380786988, "tokens_per_second": 148440.79297462487, "data_wait_s": 1.4043998817214742e-05, "gpu_peak_gb": 85.46991729736328}
 {"step": 303300, "loss": 0.5642993152141571, "lr": 0.0003, "grad_norm": 0.15055635545494098, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 44395360099, "tokens_per_second": 149010.12785942582, "data_wait_s": 1.1391999578336254e-05, "gpu_peak_gb": 85.46991729736328}
 {"step": 303400, "loss": 0.563585702329874, "lr": 0.0003, "grad_norm": 0.15217523672935845, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 44409861465, "tokens_per_second": 147336.13125779293, "data_wait_s": 1.5006000467110425e-05, "gpu_peak_gb": 85.46991729736328}
+{"step": 303500, "loss": 0.5572625398635864, "lr": 0.0003, "grad_norm": 0.14108857983182388, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 44424505103, "tokens_per_second": 149892.5680040023, "data_wait_s": 1.1225998605368659e-05, "gpu_peak_gb": 85.46991729736328}

runs/void-byte-v1/events.out.tfevents.1790703713.3e6d470a-8cc7-4d7e-a224-51a0f9a2ce55-85btc.38629.0.1790703713325746110-adad6d18

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2a19f1f442486372efa1c68fb1714497ca43317f1f9b21e3763b3dda712d3bde
-size 494673
+oid sha256:5f785f3919ba58e86f3a16f0f88f807bd43c9e9a604d9c5ae525d624d67d7993
+size 496791