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": 76830.7614319275,
-      "required_pages_per_second_upper_estimate": 0.005315330347788345
+      "required_token_positions_per_second": 76255.21795486366,
+      "required_pages_per_second_upper_estimate": 0.005275512914600214
     },
     {
       "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": 9603.845178990938,
-      "required_pages_per_second_upper_estimate": 0.004020929425350744
+      "required_token_positions_per_second": 9531.902244357958,
+      "required_pages_per_second_upper_estimate": 0.003990808420959227
     },
     {
       "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": 19207.690357981875,
-      "required_pages_per_second_upper_estimate": 0.0040004553607094874
+      "required_token_positions_per_second": 19063.804488715916,
+      "required_pages_per_second_upper_estimate": 0.0039704877286669315
     },
     {
       "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": 28811.535536972817,
-      "required_pages_per_second_upper_estimate": 0.0017272075627740135
+      "required_token_positions_per_second": 28595.706733073872,
+      "required_pages_per_second_upper_estimate": 0.0017142689555317736
     },
     {
       "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": 28811.535536972817,
-      "required_pages_per_second_upper_estimate": 0.0022399701594478954
+      "required_token_positions_per_second": 28595.706733073872,
+      "required_pages_per_second_upper_estimate": 0.0022231904192753327
     },
     {
       "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": 28811.535536972817,
-      "required_pages_per_second_upper_estimate": 0.3695634424516466
+      "required_token_positions_per_second": 28595.706733073872,
+      "required_pages_per_second_upper_estimate": 0.36679502229414546
     }
   ],
   "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": 0.4329710802967668,
-    "producer_p95_seconds": 0.2564216860009765,
-    "gpu_update_seconds": 0.9596156360538406,
-    "producer_batches_per_second": 2.3096230799400748,
-    "gpu_batches_per_second": 1.0420838952898153,
-    "producer_headroom": 2.216350420901325,
+    "producer_batch_seconds": 0.2377009172655562,
+    "producer_p95_seconds": 0.241769555999781,
+    "gpu_update_seconds": 0.9668584259196591,
+    "producer_batches_per_second": 4.206967358408692,
+    "gpu_batches_per_second": 1.0342775872784242,
+    "producer_headroom": 4.067541838046414,
     "sustainable": true,
-    "buffer_seconds": 60.45578507139196,
+    "buffer_seconds": 60.91208083293852,
     "buffer_drain_seconds": null,
     "shard_target_mib": 16,
-    "upload_pages_per_commit": 84,
+    "upload_pages_per_commit": 79,
     "max_pending_pages": 4096,
-    "generation_bytes_per_second": 341569.3922821754,
+    "generation_bytes_per_second": 340906.52654566837,
     "upload_bytes_per_second": 4968024.804064999,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 73.29284025525043,
+    "estimated_full_shards_per_hour": 73.15060469892062,
     "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

@@ -223,3 +223,4 @@
 {"step": 419400, "loss": 0.5588814616203308, "lr": 0.0003, "grad_norm": 0.18306690284279858, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 61388976968, "tokens_per_second": 151427.83731334383, "data_wait_s": 8.581999281886965e-06, "gpu_peak_gb": 89.37861251831055}
 {"step": 419500, "loss": 0.5651183724403381, "lr": 0.0003, "grad_norm": 0.15977418495314036, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 61403576810, "tokens_per_second": 150821.27804004648, "data_wait_s": 1.0966999980155379e-05, "gpu_peak_gb": 89.37861251831055}
 {"step": 419600, "loss": 0.5711472630500793, "lr": 0.0003, "grad_norm": 0.16745750840028553, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 61418193013, "tokens_per_second": 152023.9459269514, "data_wait_s": 1.3665001461049542e-05, "gpu_peak_gb": 89.37861251831055}
+{"step": 419700, "loss": 0.5668135285377502, "lr": 0.0003, "grad_norm": 0.17400848672498184, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 61432850501, "tokens_per_second": 152177.04341231627, "data_wait_s": 1.072399754775688e-05, "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 sha256:32ede230a4d0b4c6f1104c91ad08d23b9b9f68a272cc3549a91179a7bb500d10
-size 479847
+oid sha256:8f8695c3d1f8be4611ed27e2d7345dcf614287464898d5e6cb85214b3840b564
+size 481965