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": 76187.6907787204,
-      "required_pages_per_second_upper_estimate": 0.006603755940762574
+      "required_token_positions_per_second": 75565.6404015407,
+      "required_pages_per_second_upper_estimate": 0.0065498381906409055
     },
     {
       "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": 9523.46134734005,
-      "required_pages_per_second_upper_estimate": 0.003799051119889919
+      "required_token_positions_per_second": 9445.705050192588,
+      "required_pages_per_second_upper_estimate": 0.0037680329703975537
     },
     {
       "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": 19046.9226946801,
-      "required_pages_per_second_upper_estimate": 0.004406170351023807
+      "required_token_positions_per_second": 18891.410100385176,
+      "required_pages_per_second_upper_estimate": 0.004370195249261859
     },
     {
       "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": 28570.38404202015,
-      "required_pages_per_second_upper_estimate": 0.0017397864909386262
+      "required_token_positions_per_second": 28337.115150577763,
+      "required_pages_per_second_upper_estimate": 0.001725581639317073
     },
     {
       "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": 28570.38404202015,
-      "required_pages_per_second_upper_estimate": 0.002199427251439589
+      "required_token_positions_per_second": 28337.115150577763,
+      "required_pages_per_second_upper_estimate": 0.0021814695664467343
     },
     {
       "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": 28570.38404202015,
-      "required_pages_per_second_upper_estimate": 0.35474855087748675
+      "required_token_positions_per_second": 28337.115150577763,
+      "required_pages_per_second_upper_estimate": 0.351852131946531
     }
   ],
   "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": 0.06812307336727486,
-    "producer_p95_seconds": 0.24442110199743183,
-    "gpu_update_seconds": 0.9677153782510048,
-    "producer_batches_per_second": 14.679314226013512,
-    "gpu_batches_per_second": 1.0333616913346408,
-    "producer_headroom": 14.205398118692022,
+    "producer_batch_seconds": 0.09982467328140388,
+    "producer_p95_seconds": 0.32600530499985325,
+    "gpu_update_seconds": 0.9756815347322427,
+    "producer_batches_per_second": 10.017563465307006,
+    "gpu_batches_per_second": 1.0249245931198554,
+    "producer_headroom": 9.773951696108384,
     "sustainable": true,
-    "buffer_seconds": 60.9660688298133,
+    "buffer_seconds": 60.49225515339905,
     "buffer_drain_seconds": null,
     "shard_target_mib": 16,
-    "upload_pages_per_commit": 82,
+    "upload_pages_per_commit": 84,
     "max_pending_pages": 4096,
-    "generation_bytes_per_second": 213965.5787112902,
+    "generation_bytes_per_second": 212187.1168408182,
     "upload_bytes_per_second": 2244731.68118365,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 45.912032327690405,
+    "estimated_full_shards_per_hour": 45.53041581076059,
     "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

@@ -173,3 +173,4 @@
 {"step": 297500, "loss": 0.6662709899246693, "lr": 0.0003, "grad_norm": 0.1811783066483314, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 43546561769, "tokens_per_second": 132601.77143004382, "data_wait_s": 8.721999620320275e-06, "gpu_peak_gb": 85.46991729736328}
 {"step": 297600, "loss": 0.5978391785174608, "lr": 0.0003, "grad_norm": 0.15276766862988436, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 43560986804, "tokens_per_second": 147247.8215553319, "data_wait_s": 1.0168001608690247e-05, "gpu_peak_gb": 85.46991729736328}
 {"step": 297700, "loss": 0.5460868366062641, "lr": 0.0003, "grad_norm": 0.1479280371962989, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 43575536298, "tokens_per_second": 148930.1567997591, "data_wait_s": 8.151000656653196e-06, "gpu_peak_gb": 85.46991729736328}
+{"step": 297800, "loss": 0.5744358487427235, "lr": 0.0003, "grad_norm": 0.15290834698088446, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 43590185194, "tokens_per_second": 150211.0022263145, "data_wait_s": 9.883999155135825e-06, "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:e2518ee7e35575ccd3abf03a7b9b939c3e8e54b18566c3bf55984387b3578dea
-size 373164
+oid sha256:a144913a1604bc72d1e33c2b722e01ff7c93bdce3ee2cdd980cbbccd85348085
+size 375282