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": 74821.04421709034,
-      "required_pages_per_second_upper_estimate": 0.006485298480534524
+      "required_token_positions_per_second": 73431.6468088936,
+      "required_pages_per_second_upper_estimate": 0.006364869034587568
     },
     {
       "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": 9352.630527136293,
-      "required_pages_per_second_upper_estimate": 0.0037309041515622677
+      "required_token_positions_per_second": 9178.9558511117,
+      "required_pages_per_second_upper_estimate": 0.003661622726628251
     },
     {
       "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": 18705.261054272585,
-      "required_pages_per_second_upper_estimate": 0.004327132943555029
+      "required_token_positions_per_second": 18357.9117022234,
+      "required_pages_per_second_upper_estimate": 0.004246779784098257
     },
     {
       "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": 28057.891581408876,
-      "required_pages_per_second_upper_estimate": 0.0017085783889275359
+      "required_token_positions_per_second": 27536.8675533351,
+      "required_pages_per_second_upper_estimate": 0.0016768507592196545
     },
     {
       "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": 28057.891581408876,
-      "required_pages_per_second_upper_estimate": 0.0021599741631518035
+      "required_token_positions_per_second": 27536.8675533351,
+      "required_pages_per_second_upper_estimate": 0.0021198642911838643
     },
     {
       "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": 21,
-      "required_token_positions_per_second": 28057.891581408876,
-      "required_pages_per_second_upper_estimate": 0.3483851097186247
+      "required_token_positions_per_second": 27536.8675533351,
+      "required_pages_per_second_upper_estimate": 0.3419157350451979
     }
   ],
   "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": 61,
+    "prefetch_batches": 60,
     "prefetch_capacity_limit": 512,
     "batch_memory_estimate_bytes": 688946,
-    "producer_batch_seconds": 0.212897945172017,
-    "producer_p95_seconds": 0.3363575189978292,
-    "gpu_update_seconds": 0.9853912194286021,
-    "producer_batches_per_second": 4.697086198704367,
-    "gpu_batches_per_second": 1.0148253610173927,
-    "producer_headroom": 4.6284674971025535,
+    "producer_batch_seconds": 0.43800535437472377,
+    "producer_p95_seconds": 0.33220097000230453,
+    "gpu_update_seconds": 1.0040357693717215,
+    "producer_batches_per_second": 2.2830771131269705,
+    "gpu_batches_per_second": 0.9959804525945855,
+    "producer_headroom": 2.2922910858134067,
     "sustainable": true,
-    "buffer_seconds": 60.10886438514473,
+    "buffer_seconds": 60.242146162303285,
     "buffer_drain_seconds": null,
     "shard_target_mib": 16,
-    "upload_pages_per_commit": 82,
+    "upload_pages_per_commit": 74,
     "max_pending_pages": 4096,
-    "generation_bytes_per_second": 523920.2750389143,
+    "generation_bytes_per_second": 467946.7950348603,
     "upload_bytes_per_second": 4108048.426740531,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 112.42109478354999,
+    "estimated_full_shards_per_hour": 100.41048897060735,
     "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-v1/1790485060735632541-7579de36/train.jsonl

@@ -191,3 +191,4 @@
 {"step": 104100, "loss": 0.5915816277265549, "lr": 0.0003, "grad_norm": 0.1432474375497813, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 15233434216, "tokens_per_second": 148176.27755571835, "data_wait_s": 1.3013999705435708e-05, "gpu_peak_gb": 85.46991729736328}
 {"step": 104200, "loss": 0.5905342511832714, "lr": 0.0003, "grad_norm": 0.13090020853107528, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 15248055277, "tokens_per_second": 147394.1055627227, "data_wait_s": 1.8148999515688047e-05, "gpu_peak_gb": 85.46991729736328}
 {"step": 104300, "loss": 0.6315884552896023, "lr": 0.0003, "grad_norm": 0.1437229776985725, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 15262663307, "tokens_per_second": 148103.86799581692, "data_wait_s": 1.5303001418942586e-05, "gpu_peak_gb": 85.46991729736328}
+{"step": 104400, "loss": 0.5811884552240372, "lr": 0.0003, "grad_norm": 0.1369187457982779, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 15277309694, "tokens_per_second": 147965.02447973102, "data_wait_s": 1.503100065747276e-05, "gpu_peak_gb": 85.46991729736328}

runs/void-byte-v1/events.out.tfevents.1790485060.0b09bbb2-f535-42d2-bd00-cbae7fb94f51-r2qrq.28895.0.1790485060735632541-7579de36

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:879e717dbebfb25fd6a408ccc5df6458fa0242f8f6ae65e3cd8bcce955a273c8
-size 411288
+oid sha256:a706d6fcd93544de41350c5da6a82dca7a622b3ab987ed52e8377848cd2e6bb1
+size 413406