Update TensorBoard training metrics

benchmark_cache/shard_scheduler.json

@@ -22,8 +22,8 @@
       "tokens_per_second": 598315.063954628,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 1,
-      "required_token_positions_per_second": 76296.82199350644,
-      "required_pages_per_second_upper_estimate": 0.005278391178528191
+      "required_token_positions_per_second": 74372.54069216544,
+      "required_pages_per_second_upper_estimate": 0.005145264933153651
     },
     {
       "source": "rewrite6",
@@ -36,8 +36,8 @@
       "tokens_per_second": 8533.957213384612,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 1,
-      "required_token_positions_per_second": 9537.102749188305,
-      "required_pages_per_second_upper_estimate": 0.003992985763732803
+      "required_token_positions_per_second": 9296.56758652068,
+      "required_pages_per_second_upper_estimate": 0.0038922787140692424
     },
     {
       "source": "openmath",
@@ -50,8 +50,8 @@
       "tokens_per_second": 380144.35761751566,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 1,
-      "required_token_positions_per_second": 19074.20549837661,
-      "required_pages_per_second_upper_estimate": 0.003972653984686184
+      "required_token_positions_per_second": 18593.13517304136,
+      "required_pages_per_second_upper_estimate": 0.003872459722596472
     },
     {
       "source": "ultra_qa",
@@ -64,8 +64,8 @@
       "tokens_per_second": 43742.16556775435,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 1,
-      "required_token_positions_per_second": 28611.308247564917,
-      "required_pages_per_second_upper_estimate": 0.0017152042425033786
+      "required_token_positions_per_second": 27889.702759562042,
+      "required_pages_per_second_upper_estimate": 0.0016719450953253884
     },
     {
       "source": "ultra_style",
@@ -78,8 +78,8 @@
       "tokens_per_second": 22018.86276523721,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 1,
-      "required_token_positions_per_second": 28611.308247564917,
-      "required_pages_per_second_upper_estimate": 0.0022244033684032014
+      "required_token_positions_per_second": 27889.702759562042,
+      "required_pages_per_second_upper_estimate": 0.0021683017157180804
     },
     {
       "source": "cortex",
@@ -91,9 +91,9 @@
       "read_seconds": 0.04061180899998362,
       "tokens_per_second": 910315.184299418,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
-      "estimated_pages_for_buffer": 23,
-      "required_token_positions_per_second": 28611.308247564917,
-      "required_pages_per_second_upper_estimate": 0.36699514177043546
+      "estimated_pages_for_buffer": 22,
+      "required_token_positions_per_second": 27889.702759562042,
+      "required_pages_per_second_upper_estimate": 0.35773916136994194
     }
   ],
   "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": 2.558590814491007,
   "selected": {
-    "prefetch_batches": 63,
+    "prefetch_batches": 61,
     "prefetch_capacity_limit": 512,
     "batch_memory_estimate_bytes": 704566,
-    "producer_batch_seconds": 0.14762228976564984,
-    "producer_p95_seconds": 0.31985427500330843,
-    "gpu_update_seconds": 0.96633120585645,
-    "producer_batches_per_second": 6.774044770525498,
-    "gpu_batches_per_second": 1.0348418781671338,
-    "producer_headroom": 6.545970851627484,
+    "producer_batch_seconds": 0.08761908050803413,
+    "producer_p95_seconds": 0.3178241200002958,
+    "gpu_update_seconds": 0.9913336200946361,
+    "producer_batches_per_second": 11.413039194223296,
+    "gpu_batches_per_second": 1.0087421426346224,
+    "producer_headroom": 11.31412946069135,
     "sustainable": true,
-    "buffer_seconds": 60.878865968956354,
+    "buffer_seconds": 60.4713508257728,
     "buffer_drain_seconds": null,
     "shard_target_mib": 16,
-    "upload_pages_per_commit": 79,
+    "upload_pages_per_commit": 83,
     "max_pending_pages": 4096,
-    "generation_bytes_per_second": 328716.4412489728,
+    "generation_bytes_per_second": 347766.80340224586,
     "upload_bytes_per_second": 4943911.068262191,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 70.53489616491211,
+    "estimated_full_shards_per_hour": 74.6226604132703,
     "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-v1/1790989041640695387-36510d53/train.jsonl

@@ -345,3 +345,4 @@
 {"step": 591600, "loss": 0.577265739440918, "lr": 0.0003, "grad_norm": 0.19341440023267018, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 86592303768, "tokens_per_second": 151194.54178578706, "data_wait_s": 1.3630000466946512e-05, "gpu_peak_gb": 89.37899398803711}
 {"step": 591700, "loss": 0.6085643172264099, "lr": 0.0003, "grad_norm": 0.14747630772033243, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 86606910535, "tokens_per_second": 152028.13607337573, "data_wait_s": 1.573799818288535e-05, "gpu_peak_gb": 89.37899398803711}
 {"step": 591800, "loss": 0.6143639087677002, "lr": 0.0003, "grad_norm": 0.19243285126151685, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 86621653055, "tokens_per_second": 153704.5501025626, "data_wait_s": 8.636998245492578e-06, "gpu_peak_gb": 89.37899398803711}
+{"step": 591900, "loss": 0.5977855324745178, "lr": 0.0003, "grad_norm": 0.16993496243179834, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 86636358255, "tokens_per_second": 152570.45149698132, "data_wait_s": 1.22580022434704e-05, "gpu_peak_gb": 89.37899398803711}

runs/void-byte-v1/events.out.tfevents.1790989041.fb9d4967-10f3-4fa9-9622-1d5a28f8589b-vh8db.427.0.1790989041640695387-36510d53

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a85a5238c787876e78ba35d667c114caf961819646b82e3903e7fdb9a64ba156
-size 740592
+oid sha256:b49cd3f62acd3b3696f6d3b230b0318160f70fe172643ba885d4c5fe0381b594
+size 742710