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": 76533.13019365649,
-      "required_pages_per_second_upper_estimate": 0.006633697753741042
+      "required_token_positions_per_second": 75751.04195723658,
+      "required_pages_per_second_upper_estimate": 0.0065659083275926775
     },
     {
       "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": 9566.64127420706,
-      "required_pages_per_second_upper_estimate": 0.0038162762383146085
+      "required_token_positions_per_second": 9468.880244654572,
+      "required_pages_per_second_upper_estimate": 0.0037772779019684746
     },
     {
       "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": 19133.28254841412,
-      "required_pages_per_second_upper_estimate": 0.004426148183303658
+      "required_token_positions_per_second": 18937.760489309145,
+      "required_pages_per_second_upper_estimate": 0.004380917596000435
     },
     {
       "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": 28699.923822621186,
-      "required_pages_per_second_upper_estimate": 0.001747674783934526
+      "required_token_positions_per_second": 28406.64073396372,
+      "required_pages_per_second_upper_estimate": 0.0017298153825727326
     },
     {
       "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": 28699.923822621186,
-      "required_pages_per_second_upper_estimate": 0.0022093995823393203
+      "required_token_positions_per_second": 28406.64073396372,
+      "required_pages_per_second_upper_estimate": 0.0021868218383149177
     },
     {
       "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": 28699.923822621186,
-      "required_pages_per_second_upper_estimate": 0.35635700141079485
+      "required_token_positions_per_second": 28406.64073396372,
+      "required_pages_per_second_upper_estimate": 0.35271540700502524
     }
   ],
   "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": 688946,
-    "producer_batch_seconds": 0.09834142268744017,
-    "producer_p95_seconds": 0.35334287899604533,
-    "gpu_update_seconds": 0.9633475047138605,
-    "producer_batches_per_second": 10.168655004903814,
-    "gpu_batches_per_second": 1.0380470132603148,
-    "producer_headroom": 9.795948425270199,
+    "producer_batch_seconds": 0.13039045733597732,
+    "producer_p95_seconds": 0.40213196400145534,
+    "gpu_update_seconds": 0.9732935428349798,
+    "producer_batches_per_second": 7.669272893362881,
+    "gpu_batches_per_second": 1.0274392626578313,
+    "producer_headroom": 7.464453785349434,
     "sustainable": true,
-    "buffer_seconds": 60.690892796973216,
+    "buffer_seconds": 60.34419965576875,
     "buffer_drain_seconds": null,
     "shard_target_mib": 16,
     "upload_pages_per_commit": 83,
     "max_pending_pages": 4096,
-    "generation_bytes_per_second": 362449.09757456806,
+    "generation_bytes_per_second": 326244.05507966207,
     "upload_bytes_per_second": 4822326.1971726585,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 77.77313895633489,
+    "estimated_full_shards_per_hour": 70.00437964718243,
     "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

@@ -395,3 +395,4 @@
 {"step": 194500, "loss": 0.5801782682538033, "lr": 0.0003, "grad_norm": 0.1276526351946504, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 28466968617, "tokens_per_second": 149404.84753841374, "data_wait_s": 1.2269003491383046e-05, "gpu_peak_gb": 85.46427917480469}
 {"step": 194600, "loss": 0.6777424439787865, "lr": 0.0003, "grad_norm": 0.17632340123455828, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 28481509759, "tokens_per_second": 150106.601301932, "data_wait_s": 8.789000276010484e-06, "gpu_peak_gb": 85.46427917480469}
 {"step": 194700, "loss": 0.5743269063532352, "lr": 0.0003, "grad_norm": 0.14870336587591793, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 28496125885, "tokens_per_second": 131303.34703119096, "data_wait_s": 9.789997420739383e-06, "gpu_peak_gb": 85.46427917480469}
+{"step": 194800, "loss": 0.4885460287332535, "lr": 0.0003, "grad_norm": 0.15674144452471456, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 28510869924, "tokens_per_second": 150776.8410855801, "data_wait_s": 8.791997970547527e-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 sha256:80dbc887c264ffdaea09b0e71824fcab8b1487742a6b47256dfdc0238550381b
-size 846492
+oid sha256:4a6d95ecd68dec70e7f1909d139cc3c090aa5dbdcc1529738578645d931207f6
+size 848610