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": 76644.662474337,
-      "required_pages_per_second_upper_estimate": 0.005302455590093161
+      "required_token_positions_per_second": 76720.10253747494,
+      "required_pages_per_second_upper_estimate": 0.005307674708706107
     },
     {
       "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": 9580.582809292126,
-      "required_pages_per_second_upper_estimate": 0.004011189956931369
+      "required_token_positions_per_second": 9590.012817184368,
+      "required_pages_per_second_upper_estimate": 0.004015138104314893
     },
     {
       "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": 19161.16561858425,
-      "required_pages_per_second_upper_estimate": 0.003990765484432848
+      "required_token_positions_per_second": 19180.025634368736,
+      "required_pages_per_second_upper_estimate": 0.0039946935283486935
     },
     {
       "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": 28741.748427876373,
-      "required_pages_per_second_upper_estimate": 0.0017230239321424272
+      "required_token_positions_per_second": 28770.038451553104,
+      "required_pages_per_second_upper_estimate": 0.0017247198758654823
     },
     {
       "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": 28741.748427876373,
-      "required_pages_per_second_upper_estimate": 0.0022345445186766994
+      "required_token_positions_per_second": 28770.038451553104,
+      "required_pages_per_second_upper_estimate": 0.0022367439435829017
     },
     {
       "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": 28741.748427876373,
-      "required_pages_per_second_upper_estimate": 0.3686682883477171
+      "required_token_positions_per_second": 28770.038451553104,
+      "required_pages_per_second_upper_estimate": 0.3690311623959814
     }
   ],
   "gpu_measurement": "Selected forward/backward estimate; replaced by live successful update timings.",
@@ -158,24 +158,24 @@
     "prefetch_batches": 63,
     "prefetch_capacity_limit": 512,
     "batch_memory_estimate_bytes": 704566,
-    "producer_batch_seconds": 0.4400104933984217,
-    "producer_p95_seconds": 0.144557610999982,
-    "gpu_update_seconds": 0.9619456543981313,
-    "producer_batches_per_second": 2.2726730725817434,
-    "gpu_batches_per_second": 1.03955976663326,
-    "producer_headroom": 2.186187986037657,
+    "producer_batch_seconds": 0.08749761539841927,
+    "producer_p95_seconds": 0.14486193099992306,
+    "gpu_update_seconds": 0.9609997583617226,
+    "producer_batches_per_second": 11.428882895225348,
+    "gpu_batches_per_second": 1.0405829879757345,
+    "producer_headroom": 10.983153700655983,
     "sustainable": true,
-    "buffer_seconds": 60.602576227082274,
+    "buffer_seconds": 60.54298477678852,
     "buffer_drain_seconds": null,
     "shard_target_mib": 16,
-    "upload_pages_per_commit": 69,
+    "upload_pages_per_commit": 79,
     "max_pending_pages": 4096,
-    "generation_bytes_per_second": 297905.8677084344,
+    "generation_bytes_per_second": 312709.98501329083,
     "upload_bytes_per_second": 5062848.150281854,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 63.92366431655668,
+    "estimated_full_shards_per_hour": 67.10028326796574,
     "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-v1/1790824801708454555-41033672/train.jsonl

@@ -16,3 +16,4 @@
 {"step": 398700, "loss": 0.6194034218788147, "lr": 0.0003, "grad_norm": 0.14603193163626443, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 58358359562, "tokens_per_second": 152771.0292765205, "data_wait_s": 9.671000043454114e-06, "gpu_peak_gb": 89.37861251831055}
 {"step": 398800, "loss": 0.7059101462364197, "lr": 0.0003, "grad_norm": 0.30674263162538873, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 58373137537, "tokens_per_second": 135815.1271643288, "data_wait_s": 1.0689000191632658e-05, "gpu_peak_gb": 89.37861251831055}
 {"step": 398900, "loss": 0.6563550233840942, "lr": 0.0003, "grad_norm": 0.20655283200777552, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 58387974108, "tokens_per_second": 154349.77003529714, "data_wait_s": 8.372000138479052e-06, "gpu_peak_gb": 89.37861251831055}
+{"step": 399000, "loss": 0.5636134743690491, "lr": 0.0003, "grad_norm": 0.1511337301343004, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 58402707719, "tokens_per_second": 152662.34691461586, "data_wait_s": 9.202000001096167e-06, "gpu_peak_gb": 89.37861251831055}

runs/void-byte-v1/events.out.tfevents.1790824801.afc92292-ba09-4e51-8d82-bc1983323308-sdcdh.469.0.1790824801708454555-41033672

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f8d2e14bc8cf0bf69a641564eae209c2ea771afd304b86b3a087e83b8e466804
-size 38289
+oid sha256:137e65da195a471ea7c53f058e91ee237830d3dfc00f45b728290479e02d8064
+size 40407