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": 76800.71477432967,
-      "required_pages_per_second_upper_estimate": 0.005313251650297876
+      "required_token_positions_per_second": 76067.31567245613,
+      "required_pages_per_second_upper_estimate": 0.005262513398709897
     },
     {
       "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": 9600.089346791208,
-      "required_pages_per_second_upper_estimate": 0.004019356936839412
+      "required_token_positions_per_second": 9508.414459057016,
+      "required_pages_per_second_upper_estimate": 0.003980974575734453
     },
     {
       "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": 19200.178693582417,
-      "required_pages_per_second_upper_estimate": 0.003998890879110992
+      "required_token_positions_per_second": 19016.82891811403,
+      "required_pages_per_second_upper_estimate": 0.003960703956139663
     },
     {
       "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": 28800.268040373627,
-      "required_pages_per_second_upper_estimate": 0.0017265320935573633
+      "required_token_positions_per_second": 28525.24337717105,
+      "required_pages_per_second_upper_estimate": 0.0017100447849367141
     },
     {
       "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": 28800.268040373627,
-      "required_pages_per_second_upper_estimate": 0.0022390941611477883
+      "required_token_positions_per_second": 28525.24337717105,
+      "required_pages_per_second_upper_estimate": 0.002217712203289436
     },
     {
       "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": 28800.268040373627,
-      "required_pages_per_second_upper_estimate": 0.3694189151033674
+      "estimated_pages_for_buffer": 22,
+      "required_token_positions_per_second": 28525.24337717105,
+      "required_pages_per_second_upper_estimate": 0.365891194022281
     }
   ],
   "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": 62,
     "prefetch_capacity_limit": 512,
     "batch_memory_estimate_bytes": 720750,
-    "producer_batch_seconds": 0.04947380697656456,
-    "producer_p95_seconds": 0.16945143100019777,
-    "gpu_update_seconds": 0.959991065404033,
-    "producer_batches_per_second": 20.212715800781087,
-    "gpu_batches_per_second": 1.0416763614139766,
-    "producer_headroom": 19.404026576300765,
+    "producer_batch_seconds": 0.3087193143046889,
+    "producer_p95_seconds": 0.18586260600022797,
+    "gpu_update_seconds": 0.9692467697620728,
+    "producer_batches_per_second": 3.2391883295421393,
+    "gpu_batches_per_second": 1.031728999463652,
+    "producer_headroom": 3.1395728250597235,
     "sustainable": true,
-    "buffer_seconds": 60.47943712045408,
+    "buffer_seconds": 60.093299725248514,
     "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": 469559.4034046904,
-    "upload_bytes_per_second": 4361792.550967564,
+    "generation_bytes_per_second": 422854.8313396771,
+    "upload_bytes_per_second": 4384223.916194729,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 100.7565171871713,
+    "estimated_full_shards_per_hour": 90.73480324881301,
     "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": [
@@ -442,6 +442,12 @@
       "bytes": 7430334,
       "pages": 38,
       "shards": 3
+    },
+    {
+      "seconds": 1.0616652130001967,
+      "bytes": 6074717,
+      "pages": 35,
+      "shards": 2
     }
   ],
   "cache_key": {

metrics/void-byte-v1/1791342976705245012-6808dd50/train.jsonl

@@ -51,3 +51,4 @@
 {"step": 847600, "loss": 0.5515899658203125, "lr": 3e-05, "grad_norm": 0.2993019438846067, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 124065418168, "tokens_per_second": 151187.85278245917, "data_wait_s": 1.8619000002217945e-05, "gpu_peak_gb": 89.37899398803711}
 {"step": 847700, "loss": 0.5003780126571655, "lr": 3e-05, "grad_norm": 0.29074722812008036, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 124079941444, "tokens_per_second": 151178.24556929132, "data_wait_s": 1.4152000403555576e-05, "gpu_peak_gb": 89.37899398803711}
 {"step": 847800, "loss": 0.49885714054107666, "lr": 3e-05, "grad_norm": 0.2845417904857321, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 124094580343, "tokens_per_second": 152317.1407648859, "data_wait_s": 9.455000508751255e-06, "gpu_peak_gb": 89.37899398803711}
+{"step": 847900, "loss": 0.520720362663269, "lr": 3e-05, "grad_norm": 0.29118552299196276, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 124109106350, "tokens_per_second": 151040.51174555346, "data_wait_s": 2.3415000214299653e-05, "gpu_peak_gb": 89.37899398803711}

runs/void-byte-v1/events.out.tfevents.1791342976.f74d737f-395e-43a1-ae69-439b0359d002-2qp5c.418.0.1791342976705245012-6808dd50

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:19a42f9f094a5f40dc93f984c06609c32d6ed60c74b2c9022ea2a80205ea2697
-size 113202
+oid sha256:0f09cb3a732282b8b6bac68894ee9a11eb1120a2ce6cb74af5437a641c0ce463
+size 115320