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": 76227.22570091272,
-      "required_pages_per_second_upper_estimate": 0.006607182727617861
+      "required_token_positions_per_second": 75654.89508578402,
+      "required_pages_per_second_upper_estimate": 0.006557574560457188
     },
     {
       "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": 9528.40321261409,
-      "required_pages_per_second_upper_estimate": 0.0038010225038352043
+      "required_token_positions_per_second": 9456.861885723003,
+      "required_pages_per_second_upper_estimate": 0.0037724835989001925
     },
     {
       "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": 19056.80642522818,
-      "required_pages_per_second_upper_estimate": 0.0044084567781383895
+      "required_token_positions_per_second": 18913.723771446006,
+      "required_pages_per_second_upper_estimate": 0.004375357124354584
     },
     {
       "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": 28585.209637842272,
-      "required_pages_per_second_upper_estimate": 0.0017406892919403037
+      "required_token_positions_per_second": 28370.58565716901,
+      "required_pages_per_second_upper_estimate": 0.0017276198175623031
     },
     {
       "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": 28585.209637842272,
-      "required_pages_per_second_upper_estimate": 0.0022005685668458523
+      "required_token_positions_per_second": 28370.58565716901,
+      "required_pages_per_second_upper_estimate": 0.0021840462186964206
     },
     {
       "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": 28585.209637842272,
-      "required_pages_per_second_upper_estimate": 0.3549326351595201
+      "required_token_positions_per_second": 28370.58565716901,
+      "required_pages_per_second_upper_estimate": 0.3522677236198146
     }
   ],
   "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": 0.138298248523256,
-    "producer_p95_seconds": 0.3617938450042857,
-    "gpu_update_seconds": 0.9672134768393807,
-    "producer_batches_per_second": 7.230749562470718,
-    "gpu_batches_per_second": 1.0338979180353831,
-    "producer_headroom": 6.993678424472134,
+    "producer_batch_seconds": 0.20794020268755276,
+    "producer_p95_seconds": 0.3620124630033388,
+    "gpu_update_seconds": 0.9745304638437586,
+    "producer_batches_per_second": 4.809074854575294,
+    "gpu_batches_per_second": 1.02613518725293,
+    "producer_headroom": 4.686589948688618,
     "sustainable": true,
-    "buffer_seconds": 60.934449040880985,
+    "buffer_seconds": 60.42088875831303,
     "buffer_drain_seconds": null,
     "shard_target_mib": 16,
-    "upload_pages_per_commit": 82,
+    "upload_pages_per_commit": 83,
     "max_pending_pages": 4096,
-    "generation_bytes_per_second": 402538.75574420334,
+    "generation_bytes_per_second": 472959.3678470888,
     "upload_bytes_per_second": 5568279.627344222,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 86.3754463600595,
+    "estimated_full_shards_per_hour": 101.48607040938852,
     "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

@@ -299,3 +299,4 @@
 {"step": 310100, "loss": 0.6318192221224308, "lr": 0.0003, "grad_norm": 0.19374741506262244, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 45390295004, "tokens_per_second": 136630.82280010704, "data_wait_s": 9.717005013953894e-06, "gpu_peak_gb": 85.46991729736328}
 {"step": 310200, "loss": 0.5802470557391644, "lr": 0.0003, "grad_norm": 0.14023435505623016, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 45404959745, "tokens_per_second": 150899.77110123532, "data_wait_s": 8.891001925803721e-06, "gpu_peak_gb": 85.46991729736328}
 {"step": 310300, "loss": 0.6191430687904358, "lr": 0.0003, "grad_norm": 0.14687646478434319, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 45419718011, "tokens_per_second": 151917.81851201394, "data_wait_s": 1.1657997674774379e-05, "gpu_peak_gb": 85.46991729736328}
+{"step": 310400, "loss": 0.607383145019412, "lr": 0.0003, "grad_norm": 0.17157508109089079, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 45434509296, "tokens_per_second": 151587.65096832925, "data_wait_s": 1.2898999557364732e-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:fdd9ae349d6e50511e39f8b4a447b60ccb3454b7c0a8f20a3cb1252993b457aa
-size 642381
+oid sha256:d789edb1f3916ac3a7d9ec272559a136ab3d2bdf04afb87a6934f21364ea9c95
+size 644499