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": 77457.49977768447,
-      "required_pages_per_second_upper_estimate": 0.005358689560781129
+      "required_token_positions_per_second": 77275.73313217435,
+      "required_pages_per_second_upper_estimate": 0.0053461145224880146
     },
     {
       "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": 9682.187472210559,
-      "required_pages_per_second_upper_estimate": 0.0040537297075486835
+      "required_token_positions_per_second": 9659.466641521794,
+      "required_pages_per_second_upper_estimate": 0.004044216970204195
     },
     {
       "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": 19364.374944421117,
-      "required_pages_per_second_upper_estimate": 0.004033088628014369
+      "required_token_positions_per_second": 19318.93328304359,
+      "required_pages_per_second_upper_estimate": 0.004023624328326627
     },
     {
       "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": 29046.562416631677,
-      "required_pages_per_second_upper_estimate": 0.0017412970653442959
+      "required_token_positions_per_second": 28978.39992456538,
+      "required_pages_per_second_upper_estimate": 0.0017372108280230205
     },
     {
       "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": 29046.562416631677,
-      "required_pages_per_second_upper_estimate": 0.002258242465567381
+      "required_token_positions_per_second": 28978.39992456538,
+      "required_pages_per_second_upper_estimate": 0.0022529431316243437
     },
     {
       "source": "cortex",
@@ -92,8 +92,8 @@
       "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": 29046.562416631677,
-      "required_pages_per_second_upper_estimate": 0.3725781149117081
+      "required_token_positions_per_second": 28978.39992456538,
+      "required_pages_per_second_upper_estimate": 0.3717037996506635
     }
   ],
   "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": 64,
+    "prefetch_batches": 63,
     "prefetch_capacity_limit": 512,
     "batch_memory_estimate_bytes": 720750,
-    "producer_batch_seconds": 0.052001442203163606,
-    "producer_p95_seconds": 0.17696672000056424,
-    "gpu_update_seconds": 0.9518510178047479,
-    "producer_batches_per_second": 19.230235886403225,
-    "gpu_batches_per_second": 1.0505845781478471,
-    "producer_headroom": 18.3043196010983,
+    "producer_batch_seconds": 0.09133185899220564,
+    "producer_p95_seconds": 0.19984989699878497,
+    "gpu_update_seconds": 0.9540899453376105,
+    "producer_batches_per_second": 10.949081854179067,
+    "gpu_batches_per_second": 1.0481192102345696,
+    "producer_headroom": 10.44640890775073,
     "sustainable": true,
-    "buffer_seconds": 60.918465139503866,
+    "buffer_seconds": 60.107666556269464,
     "buffer_drain_seconds": null,
     "shard_target_mib": 16,
-    "upload_pages_per_commit": 84,
+    "upload_pages_per_commit": 82,
     "max_pending_pages": 4096,
-    "generation_bytes_per_second": 347423.52919589734,
+    "generation_bytes_per_second": 350086.9373298794,
     "upload_bytes_per_second": 3053438.887152815,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 74.54900175960245,
+    "estimated_full_shards_per_hour": 75.12050714418685,
     "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-v1/1791083219122319327-61ae60ac/train.jsonl

@@ -82,3 +82,4 @@
 {"step": 653200, "loss": 0.5545850396156311, "lr": 0.0001, "grad_norm": 0.21726353145859434, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 95611313773, "tokens_per_second": 153295.92390513566, "data_wait_s": 7.992999599082395e-06, "gpu_peak_gb": 89.37899398803711}
 {"step": 653300, "loss": 0.5329073667526245, "lr": 0.0001, "grad_norm": 0.18291522439644012, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 95626076045, "tokens_per_second": 154234.73919961223, "data_wait_s": 9.192999641527422e-06, "gpu_peak_gb": 89.37899398803711}
 {"step": 653400, "loss": 0.5343234539031982, "lr": 0.0001, "grad_norm": 0.17442434352286854, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 95640817371, "tokens_per_second": 153838.3333628902, "data_wait_s": 1.3274000593810342e-05, "gpu_peak_gb": 89.37899398803711}
+{"step": 653500, "loss": 0.5341771841049194, "lr": 0.0001, "grad_norm": 0.2140702360987653, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 95655384593, "tokens_per_second": 151370.2923929644, "data_wait_s": 1.1362000805092975e-05, "gpu_peak_gb": 89.37899398803711}

runs/void-byte-v1/events.out.tfevents.1791083219.1a73f132-ede3-4b51-9e4f-0cafff989449-8p4r7.435.0.1791083219122319327-61ae60ac

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1a9df700e4dbdf3fdf1abcb04f5e10f3e9ac698fd2b44050d4ad22e0d765b31d
-size 178860
+oid sha256:769036763d767e66a4d8147026a406d56b9b976455dc007cb873ffe4dd478c5e
+size 180978