Nemotron-3.5-Parse / reference /validation /batch_benchmarks.json
emelryan's picture sosahi's picture ronay-nv's picture
Nemotron 3.5 Parse release
dca5b2d
Raw History Blame Contribute Delete
37.4 kB
{
"complete": true,
"purpose": "Prepared-page single/double/triple throughput and staged versus pipelined scheduling; guarded refinement avoids overlapping same-class crop additions.",
"package_version": "3.5.0.post4",
"gpu_candidate_wheel_sha256": "e7a3eacb6bbcf3c60c6034e45b1cd3fc16e90ba4ab0af2616c6b349044bf174e",
"release_wheel_equivalence": {
"gpu_candidate_wheel_sha256": "e7a3eacb6bbcf3c60c6034e45b1cd3fc16e90ba4ab0af2616c6b349044bf174e",
"release_wheel_sha256": "09eb1c41230adeb5d0272e4f4ed741288f2d67086f10ed6d56cbe0601de8fcc4",
"prepared_batch_runtime_python_ast_equal": true,
"other_runtime_ast_equal_except_parse_many_default_resolution": true,
"python_modules_checked": 32,
"changed_archive_members": [
"nemotron_parse/SOURCE_MANIFEST.json",
"nemotron_parse/inference.py",
"nemotron_parse-3.5.0.post4.dist-info/RECORD"
],
"change_scope": "Prepared-batch benchmark path unchanged. parse_many now resolves its default to 128 HTTP requests or 4 custom-backend workers; batching docstring and non-executable source provenance corrected. Final installed wheel passed 74 tests including the HTTP default concurrency check."
},
"release_model_sha256": "81dd7609c589f3694ce75520c25af5010953231358de6aec4d03e8c6aac787e9",
"hardware": "one H100 80GB per run",
"server": {
"vllm": "0.26.0",
"max_num_seqs": 192,
"max_num_batched_tokens": 16384,
"max_model_len": 16384,
"one_token_mtp": true,
"distributed_executor_backend": "uni",
"tensor_parallel_size": 1,
"prefix_caching": false,
"flash_attention_version": 3
},
"runtime_versions": {
"torch": "2.11.0+cu130",
"torchvision": "0.26.0+cu130",
"transformers": "5.14.1",
"huggingface-hub": "1.24.0",
"accelerate": "1.12.0",
"timm": "1.0.22",
"einops": "0.8.1",
"Pillow": "12.2.0",
"numpy": "2.2.6",
"safetensors": "0.8.0",
"beautifulsoup4": "4.13.3",
"vllm": "0.26.0",
"openai": "2.30.0",
"httpx": "0.28.1",
"nemotron-parse": "3.5.0.post4"
},
"sampling": {
"seed": 20261012,
"top_k": 50,
"top_p": 1.0,
"fullpage": {
"temperature": 0.2,
"repetition_penalty": 1.05,
"max_tokens": 8744
},
"detection": {
"temperature": 0.2,
"repetition_penalty": 1.05,
"max_tokens": 4096
},
"crop": {
"temperature": 0.1,
"repetition_penalty": 1.0,
"max_tokens": 8744
}
},
"preparation": {
"pages": 2078,
"preprocessing_seconds": 9.701208557002246,
"decoded_pixels_and_encoded_requests_cached": true
},
"concurrency_selection": {
"complete": true,
"selected_concurrency": {
"double": 256,
"triple": 128
},
"scope": "256-page pipelined sample"
},
"runs": {
"double_pipelined": {
"mode": "double",
"schedule": "pipelined",
"peak_inflight_requests": 256,
"request_service_seconds": {
"fullpage": 28512.122418867133,
"crop": 13850.49556360254
},
"stage_first_request_seconds": {
"fullpage": 0.0008634909754619002,
"crop": 8.29464137996547
},
"overlapping_stage_times": true,
"pages": 2078,
"complete_pages": 2038,
"incomplete_pages": 40,
"concurrency": 256,
"postprocess_workers": 8,
"preprocessing_seconds": 9.701208557002246,
"pipeline_seconds": 175.00423502299236,
"processed_pages_per_second": 11.87399836196529,
"complete_pages_per_second": 11.645432464718604,
"stage_times": {
"fullpage_span_seconds": 174.67124180600513,
"crop_span_seconds": 166.61445210501552
},
"requests": {
"fullpage": 2078,
"detection": 0,
"crop": 2331
},
"token_limited_responses": {
"fullpage": 33,
"crop": 8
},
"completion_tokens": 3832237,
"postprocessing": {
"infer_headers": true,
"render_charts": true,
"split_chart_value_percent": true,
"relocate_chart_captions": true,
"normalize_text": true,
"deduplicate": true
},
"allow_partial": true,
"timing_excludes": [
"initial image preparation",
"output serialization",
"server startup",
"warmup requests"
],
"timing_includes": [
"all mode inference stages",
"predicted-region crop preparation",
"merge and document rendering"
],
"keys_sha256": "e2800a56d79085a612e2677e87fabd9735be7d92479cd68ad750ff3b49f67d3a",
"vllm_queue": {
"sample_interval_seconds": 0.5,
"samples": 344,
"peak_running": 192.0,
"peak_waiting": 64.0,
"fraction_samples_with_waiting": 0.9215116279069767,
"mean_running": 178.1656976744186
},
"output_serialization_seconds": 1.3962035359581932,
"pipeline_plus_serialization_seconds": 176.40043855895055,
"pages_per_second_with_serialization": 11.780016064447377,
"verified_unique_prediction_rows": 2078
},
"double_staged": {
"mode": "double",
"schedule": "staged",
"pages": 2078,
"complete_pages": 2046,
"incomplete_pages": 32,
"concurrency": 256,
"postprocess_workers": 8,
"preprocessing_seconds": 9.701208557002246,
"pipeline_seconds": 186.15673828299623,
"processed_pages_per_second": 11.162636492056581,
"complete_pages_per_second": 10.990738336259751,
"stage_times": {
"fullpage_seconds": 135.18058688100427,
"crop_preprocessing_seconds": 3.983198474976234,
"crop_seconds": 43.71528333501192,
"postprocessing_seconds": 3.214057782024611
},
"requests": {
"fullpage": 2078,
"detection": 0,
"crop": 2345
},
"token_limited_responses": {
"crop": 5,
"fullpage": 28
},
"completion_tokens": 3798721,
"postprocessing": {
"infer_headers": true,
"render_charts": true,
"split_chart_value_percent": true,
"relocate_chart_captions": true,
"normalize_text": true,
"deduplicate": true
},
"allow_partial": true,
"timing_excludes": [
"initial image preparation",
"output serialization",
"server startup",
"warmup requests"
],
"timing_includes": [
"all mode inference stages",
"predicted-region crop preparation",
"merge and document rendering"
],
"keys_sha256": "e2800a56d79085a612e2677e87fabd9735be7d92479cd68ad750ff3b49f67d3a",
"vllm_queue": {
"sample_interval_seconds": 0.5,
"samples": 365,
"peak_running": 192.0,
"peak_waiting": 64.0,
"fraction_samples_with_waiting": 0.8301369863013699,
"mean_running": 163.14794520547946
},
"output_serialization_seconds": 1.5585761109832674,
"pipeline_plus_serialization_seconds": 187.7153143939795,
"pages_per_second_with_serialization": 11.069954557030252,
"verified_unique_prediction_rows": 2078
},
"single_pipelined": {
"mode": "single",
"schedule": "pipelined",
"peak_inflight_requests": 128,
"request_service_seconds": {
"fullpage": 16892.07212134439
},
"stage_first_request_seconds": {
"fullpage": 0.000805519986897707
},
"overlapping_stage_times": true,
"pages": 2078,
"complete_pages": 2047,
"incomplete_pages": 31,
"concurrency": 128,
"postprocess_workers": 8,
"preprocessing_seconds": 9.701208557002246,
"pipeline_seconds": 139.41391940996982,
"processed_pages_per_second": 14.905254861168455,
"complete_pages_per_second": 14.682895428687116,
"stage_times": {
"fullpage_span_seconds": 139.2903347380343
},
"requests": {
"fullpage": 2078,
"detection": 0,
"crop": 0
},
"token_limited_responses": {
"fullpage": 31
},
"completion_tokens": 2695113,
"postprocessing": {
"infer_headers": true,
"render_charts": true,
"split_chart_value_percent": true,
"relocate_chart_captions": true,
"normalize_text": true,
"deduplicate": true
},
"allow_partial": true,
"timing_excludes": [
"initial image preparation",
"output serialization",
"server startup",
"warmup requests"
],
"timing_includes": [
"all mode inference stages",
"predicted-region crop preparation",
"merge and document rendering"
],
"keys_sha256": "e2800a56d79085a612e2677e87fabd9735be7d92479cd68ad750ff3b49f67d3a",
"vllm_queue": {
"sample_interval_seconds": 0.5,
"samples": 275,
"peak_running": 128.0,
"peak_waiting": 9.0,
"fraction_samples_with_waiting": 0.01818181818181818,
"mean_running": 114.3709090909091
},
"output_serialization_seconds": 1.4439532819669694,
"pipeline_plus_serialization_seconds": 140.8578726919368,
"pages_per_second_with_serialization": 14.752459058818031,
"verified_unique_prediction_rows": 2078
},
"single_staged": {
"mode": "single",
"schedule": "staged",
"pages": 2078,
"complete_pages": 2048,
"incomplete_pages": 30,
"concurrency": 128,
"postprocess_workers": 8,
"preprocessing_seconds": 9.701208557002246,
"pipeline_seconds": 142.67479116702452,
"processed_pages_per_second": 14.564591144677802,
"complete_pages_per_second": 14.354322745091501,
"stage_times": {
"fullpage_seconds": 139.43862492003245,
"postprocessing_seconds": 3.2128840790246613
},
"requests": {
"fullpage": 2078,
"detection": 0,
"crop": 0
},
"token_limited_responses": {
"fullpage": 30
},
"completion_tokens": 2695641,
"postprocessing": {
"infer_headers": true,
"render_charts": true,
"split_chart_value_percent": true,
"relocate_chart_captions": true,
"normalize_text": true,
"deduplicate": true
},
"allow_partial": true,
"timing_excludes": [
"initial image preparation",
"output serialization",
"server startup",
"warmup requests"
],
"timing_includes": [
"all mode inference stages",
"predicted-region crop preparation",
"merge and document rendering"
],
"keys_sha256": "e2800a56d79085a612e2677e87fabd9735be7d92479cd68ad750ff3b49f67d3a",
"vllm_queue": {
"sample_interval_seconds": 0.5,
"samples": 281,
"peak_running": 128.0,
"peak_waiting": 10.0,
"fraction_samples_with_waiting": 0.021352313167259787,
"mean_running": 111.33807829181495
},
"output_serialization_seconds": 1.3698211470036767,
"pipeline_plus_serialization_seconds": 144.0446123140282,
"pages_per_second_with_serialization": 14.42608624243302,
"verified_unique_prediction_rows": 2078
},
"triple_pipelined": {
"mode": "triple",
"schedule": "pipelined",
"peak_inflight_requests": 128,
"request_service_seconds": {
"detection": 2041.2865145394462,
"crop": 6923.074506564648,
"fullpage": 14563.87797958206
},
"stage_first_request_seconds": {
"detection": 0.002384984982199967,
"crop": 1.4474368169903755,
"fullpage": 0.00044989498564973474
},
"overlapping_stage_times": true,
"pages": 2078,
"complete_pages": 2039,
"incomplete_pages": 39,
"concurrency": 128,
"postprocess_workers": 8,
"preprocessing_seconds": 9.701208557002246,
"pipeline_seconds": 190.63594810198992,
"processed_pages_per_second": 10.900357569959855,
"complete_pages_per_second": 10.695779155509213,
"stage_times": {
"detection_span_seconds": 183.4038801449933,
"crop_span_seconds": 189.06292444997234,
"fullpage_span_seconds": 190.37763873301446
},
"requests": {
"fullpage": 2078,
"detection": 2078,
"crop": 2610
},
"token_limited_responses": {
"fullpage": 31,
"crop": 22
},
"completion_tokens": 4240052,
"postprocessing": {
"infer_headers": true,
"render_charts": true,
"split_chart_value_percent": true,
"relocate_chart_captions": true,
"normalize_text": true,
"deduplicate": true
},
"allow_partial": true,
"timing_excludes": [
"initial image preparation",
"output serialization",
"server startup",
"warmup requests"
],
"timing_includes": [
"all mode inference stages",
"predicted-region crop preparation",
"merge and document rendering"
],
"keys_sha256": "e2800a56d79085a612e2677e87fabd9735be7d92479cd68ad750ff3b49f67d3a",
"vllm_queue": {
"sample_interval_seconds": 0.5,
"samples": 375,
"peak_running": 128.0,
"peak_waiting": 4.0,
"fraction_samples_with_waiting": 0.013333333333333334,
"mean_running": 115.208
},
"output_serialization_seconds": 1.5289277550182305,
"pipeline_plus_serialization_seconds": 192.16487585700816,
"pages_per_second_with_serialization": 10.813630694645056,
"verified_unique_prediction_rows": 2078
},
"triple_staged": {
"mode": "triple",
"schedule": "staged",
"pages": 2078,
"complete_pages": 2043,
"incomplete_pages": 35,
"concurrency": 128,
"postprocess_workers": 8,
"preprocessing_seconds": 9.701208557002246,
"pipeline_seconds": 284.1452879360295,
"processed_pages_per_second": 7.313160162162628,
"complete_pages_per_second": 7.189983739797039,
"stage_times": {
"fullpage_seconds": 137.60017299599713,
"detection_seconds": 89.23946119600441,
"crop_preprocessing_seconds": 4.473296172975097,
"crop_seconds": 49.61852987698512,
"postprocessing_seconds": 3.1865693450090475
},
"requests": {
"fullpage": 2078,
"detection": 2078,
"crop": 2611
},
"token_limited_responses": {
"fullpage": 26,
"crop": 23
},
"completion_tokens": 4214735,
"postprocessing": {
"infer_headers": true,
"render_charts": true,
"split_chart_value_percent": true,
"relocate_chart_captions": true,
"normalize_text": true,
"deduplicate": true
},
"allow_partial": true,
"timing_excludes": [
"initial image preparation",
"output serialization",
"server startup",
"warmup requests"
],
"timing_includes": [
"all mode inference stages",
"predicted-region crop preparation",
"merge and document rendering"
],
"keys_sha256": "e2800a56d79085a612e2677e87fabd9735be7d92479cd68ad750ff3b49f67d3a",
"vllm_queue": {
"sample_interval_seconds": 0.5,
"samples": 557,
"peak_running": 128.0,
"peak_waiting": 36.0,
"fraction_samples_with_waiting": 0.1687612208258528,
"mean_running": 97.89766606822262
},
"output_serialization_seconds": 1.5726757429656573,
"pipeline_plus_serialization_seconds": 285.71796367899515,
"pages_per_second_with_serialization": 7.272906376774539,
"verified_unique_prediction_rows": 2078
}
},
"earlier_staged_runs": {
"double": {
"mode": "double",
"pages": 2078,
"complete_pages": 2045,
"incomplete_pages": 33,
"concurrency": 256,
"postprocess_workers": 8,
"preprocessing_seconds": 10.932252227998106,
"pipeline_seconds": 191.57418219500687,
"processed_pages_per_second": 10.846973095178168,
"complete_pages_per_second": 10.674716063349063,
"stage_times": {
"fullpage_seconds": 134.28128188298433,
"crop_preprocessing_seconds": 4.702835199015681,
"crop_seconds": 49.023790333012585,
"postprocessing_seconds": 3.4980894689797424
},
"requests": {
"fullpage": 2078,
"detection": 0,
"crop": 2330
},
"token_limited_responses": {
"fullpage": 29,
"crop": 5
},
"completion_tokens": 3783551,
"postprocessing": {
"infer_headers": true,
"render_charts": true,
"split_chart_value_percent": true,
"relocate_chart_captions": true,
"normalize_text": true,
"deduplicate": true
},
"allow_partial": true,
"timing_excludes": [
"initial image preparation",
"output serialization",
"server startup",
"warmup requests"
],
"timing_includes": [
"all mode inference stages",
"predicted-region crop preparation",
"merge and document rendering"
],
"keys_sha256": "e2800a56d79085a612e2677e87fabd9735be7d92479cd68ad750ff3b49f67d3a",
"output_serialization_seconds": 1.5591519760200754,
"pipeline_plus_serialization_seconds": 193.13333417102695,
"pages_per_second_with_serialization": 10.759406235693376,
"verified_unique_prediction_rows": 2078
},
"single_c128": {
"mode": "single",
"pages": 2078,
"complete_pages": 2052,
"incomplete_pages": 26,
"concurrency": 128,
"postprocess_workers": 8,
"preprocessing_seconds": 10.932252227998106,
"pipeline_seconds": 174.87433017898002,
"processed_pages_per_second": 11.882818924156638,
"complete_pages_per_second": 11.734140727800492,
"stage_times": {
"fullpage_seconds": 171.50884548400063,
"postprocessing_seconds": 3.302743566018762
},
"requests": {
"fullpage": 2078,
"detection": 0,
"crop": 0
},
"token_limited_responses": {
"fullpage": 26
},
"completion_tokens": 2680888,
"postprocessing": {
"infer_headers": true,
"render_charts": true,
"split_chart_value_percent": true,
"relocate_chart_captions": true,
"normalize_text": true,
"deduplicate": true
},
"allow_partial": true,
"timing_excludes": [
"initial image preparation",
"output serialization",
"server startup",
"warmup requests"
],
"timing_includes": [
"all mode inference stages",
"predicted-region crop preparation",
"merge and document rendering"
],
"keys_sha256": "e2800a56d79085a612e2677e87fabd9735be7d92479cd68ad750ff3b49f67d3a",
"output_serialization_seconds": 1.3440068789932411,
"pipeline_plus_serialization_seconds": 176.21833705797326,
"pages_per_second_with_serialization": 11.792189364018164,
"verified_unique_prediction_rows": 2078
},
"single_c256": {
"mode": "single",
"pages": 2078,
"complete_pages": 2045,
"incomplete_pages": 33,
"concurrency": 256,
"postprocess_workers": 8,
"preprocessing_seconds": 10.932252227998106,
"pipeline_seconds": 137.7506838600093,
"processed_pages_per_second": 15.08522456492333,
"complete_pages_per_second": 14.845661325923103,
"stage_times": {
"fullpage_seconds": 134.78283468799782,
"postprocessing_seconds": 2.9462760030000936
},
"requests": {
"fullpage": 2078,
"detection": 0,
"crop": 0
},
"token_limited_responses": {
"fullpage": 33
},
"completion_tokens": 2719778,
"postprocessing": {
"infer_headers": true,
"render_charts": true,
"split_chart_value_percent": true,
"relocate_chart_captions": true,
"normalize_text": true,
"deduplicate": true
},
"allow_partial": true,
"timing_excludes": [
"initial image preparation",
"output serialization",
"server startup",
"warmup requests"
],
"timing_includes": [
"all mode inference stages",
"predicted-region crop preparation",
"merge and document rendering"
],
"keys_sha256": "e2800a56d79085a612e2677e87fabd9735be7d92479cd68ad750ff3b49f67d3a",
"output_serialization_seconds": 1.2913870430202223,
"pipeline_plus_serialization_seconds": 139.04207090302953,
"pages_per_second_with_serialization": 14.945116873649235,
"verified_unique_prediction_rows": 2078
},
"triple": {
"mode": "triple",
"pages": 2078,
"complete_pages": 2045,
"incomplete_pages": 33,
"concurrency": 256,
"postprocess_workers": 8,
"preprocessing_seconds": 10.932252227998106,
"pipeline_seconds": 282.9865308199951,
"processed_pages_per_second": 7.343105673541031,
"complete_pages_per_second": 7.226492349562758,
"stage_times": {
"fullpage_seconds": 134.07296904400573,
"detection_seconds": 86.69076943298569,
"crop_preprocessing_seconds": 5.352364441001555,
"crop_seconds": 53.1396865429997,
"postprocessing_seconds": 3.6966218999878038
},
"requests": {
"fullpage": 2078,
"detection": 2078,
"crop": 2619
},
"token_limited_responses": {
"fullpage": 28,
"crop": 22
},
"completion_tokens": 4206354,
"postprocessing": {
"infer_headers": true,
"render_charts": true,
"split_chart_value_percent": true,
"relocate_chart_captions": true,
"normalize_text": true,
"deduplicate": true
},
"allow_partial": true,
"timing_excludes": [
"initial image preparation",
"output serialization",
"server startup",
"warmup requests"
],
"timing_includes": [
"all mode inference stages",
"predicted-region crop preparation",
"merge and document rendering"
],
"keys_sha256": "e2800a56d79085a612e2677e87fabd9735be7d92479cd68ad750ff3b49f67d3a",
"output_serialization_seconds": 1.8944782140024472,
"pipeline_plus_serialization_seconds": 284.88100903399754,
"pages_per_second_with_serialization": 7.294273518077903,
"verified_unique_prediction_rows": 2078
}
},
"comparison": {
"single": {
"concurrency": 128,
"staged_pages_per_second": 14.564591144677802,
"pipelined_pages_per_second": 14.905254861168455,
"throughput_ratio": 1.0233898578481648,
"elapsed_reduction_percent": 2.285527618706873,
"staged_completion_tokens": 2695641,
"pipelined_completion_tokens": 2695113,
"staged_complete_pages": 2048,
"pipelined_complete_pages": 2047
},
"double": {
"concurrency": 256,
"staged_pages_per_second": 11.162636492056581,
"pipelined_pages_per_second": 11.87399836196529,
"throughput_ratio": 1.063727047854234,
"elapsed_reduction_percent": 5.990921071602462,
"staged_completion_tokens": 3798721,
"pipelined_completion_tokens": 3832237,
"staged_complete_pages": 2046,
"pipelined_complete_pages": 2038
},
"triple": {
"concurrency": 128,
"staged_pages_per_second": 7.313160162162628,
"pipelined_pages_per_second": 10.900357569959855,
"throughput_ratio": 1.4905126276813867,
"elapsed_reduction_percent": 32.90898839578561,
"staged_completion_tokens": 4214735,
"pipelined_completion_tokens": 4240052,
"staged_complete_pages": 2043,
"pipelined_complete_pages": 2039
}
},
"official_quality": {
"baseline/double": {
"complete": true,
"pages": 2078,
"fresh_page_evaluations": 2078,
"reused_page_scores": 0,
"official_errors": 0,
"scores": {
"charts": 0.577349577261549,
"visual_grounding": 0.7053143510345445,
"tables": 0.8169993270165995,
"content_faithfulness": 0.8814658795594027,
"semantic_formatting": 0.6802102726893035,
"overall": 0.7322678815122798
},
"independent_scores": {
"tables": 0.8169993270165995,
"charts": 0.577349577261549,
"content_faithfulness": 0.8814658795594027,
"semantic_formatting": 0.6802102726893035,
"visual_grounding": 0.7053143510345445,
"overall": 0.7322678815122798
},
"denominators": {
"tables": 503,
"charts": 568,
"content_faithfulness": 506,
"semantic_formatting": 456,
"visual_grounding": 500
},
"prediction_sha256": "925036f73d20d0ebf7a3411a10f662d4996c8d4ac93a2f30db02547bbe431524",
"seconds": 112.09945668200089
},
"baseline/single": {
"complete": true,
"pages": 2078,
"fresh_page_evaluations": 2078,
"reused_page_scores": 0,
"official_errors": 0,
"scores": {
"charts": 0.564825887141028,
"visual_grounding": 0.7048062312678962,
"tables": 0.8058640339417577,
"content_faithfulness": 0.8806442186073474,
"semantic_formatting": 0.6806519653753794,
"overall": 0.7273584672666817
},
"independent_scores": {
"tables": 0.8058640339417577,
"charts": 0.564825887141028,
"content_faithfulness": 0.8806442186073474,
"semantic_formatting": 0.6806519653753794,
"visual_grounding": 0.7048062312678962,
"overall": 0.7273584672666817
},
"denominators": {
"tables": 503,
"charts": 568,
"content_faithfulness": 506,
"semantic_formatting": 456,
"visual_grounding": 500
},
"prediction_sha256": "8a139e898d71fd206340887045cff80b0ba2597e96910057002ad4fe35302e4e",
"seconds": 108.71837869426236
},
"baseline/triple": {
"complete": true,
"pages": 2078,
"fresh_page_evaluations": 2078,
"reused_page_scores": 0,
"official_errors": 0,
"scores": {
"charts": 0.5819361878340752,
"visual_grounding": 0.7047845094496455,
"tables": 0.8234070433681294,
"content_faithfulness": 0.8808115694155612,
"semantic_formatting": 0.6812589701148343,
"overall": 0.7344396560364491
},
"independent_scores": {
"tables": 0.8234070433681294,
"charts": 0.5819361878340752,
"content_faithfulness": 0.8808115694155612,
"semantic_formatting": 0.6812589701148343,
"visual_grounding": 0.7047845094496455,
"overall": 0.7344396560364491
},
"denominators": {
"tables": 503,
"charts": 568,
"content_faithfulness": 506,
"semantic_formatting": 456,
"visual_grounding": 500
},
"prediction_sha256": "a849a739ddeebce793d8f9684b9141a7dd29b95d4d74d88bd8f5ce445866f55b",
"seconds": 111.65782342711464
},
"overlap/double_pipelined": {
"complete": true,
"pages": 2078,
"fresh_page_evaluations": 2078,
"reused_page_scores": 0,
"official_errors": 0,
"scores": {
"charts": 0.5786840995467756,
"visual_grounding": 0.7049578809213192,
"tables": 0.8139896349961072,
"content_faithfulness": 0.8808529188100122,
"semantic_formatting": 0.6857629648330742,
"overall": 0.7328494998214576
},
"independent_scores": {
"tables": 0.8139896349961072,
"charts": 0.5786840995467756,
"content_faithfulness": 0.8808529188100122,
"semantic_formatting": 0.6857629648330742,
"visual_grounding": 0.7049578809213192,
"overall": 0.7328494998214576
},
"denominators": {
"tables": 503,
"charts": 568,
"content_faithfulness": 506,
"semantic_formatting": 456,
"visual_grounding": 500
},
"prediction_sha256": "089afb763fa794efd9e11e29e69a2dc833d350814eea09b4e09e6f9c8c9a8706",
"seconds": 113.78007110398903
},
"overlap/double_staged": {
"complete": true,
"pages": 2078,
"fresh_page_evaluations": 2078,
"reused_page_scores": 0,
"official_errors": 0,
"scores": {
"charts": 0.5782609800215434,
"visual_grounding": 0.7044372959382412,
"tables": 0.8177032215627547,
"content_faithfulness": 0.8812009445647999,
"semantic_formatting": 0.6806841200304778,
"overall": 0.7324573124235634
},
"independent_scores": {
"tables": 0.8177032215627547,
"charts": 0.5782609800215434,
"content_faithfulness": 0.8812009445647999,
"semantic_formatting": 0.6806841200304778,
"visual_grounding": 0.7044372959382412,
"overall": 0.7324573124235634
},
"denominators": {
"tables": 503,
"charts": 568,
"content_faithfulness": 506,
"semantic_formatting": 456,
"visual_grounding": 500
},
"prediction_sha256": "8a58fee69060853d9790cb4b3035ad2871b8d80087cb1a743d9e41da1cf2052b",
"seconds": 111.54542283300543
},
"overlap/single_pipelined": {
"complete": true,
"pages": 2078,
"fresh_page_evaluations": 2078,
"reused_page_scores": 0,
"official_errors": 0,
"scores": {
"charts": 0.5673926767676768,
"visual_grounding": 0.7045478807517503,
"tables": 0.8040366382462675,
"content_faithfulness": 0.8813186137735038,
"semantic_formatting": 0.6821703978529401,
"overall": 0.7278932414784277
},
"independent_scores": {
"tables": 0.8040366382462675,
"charts": 0.5673926767676768,
"content_faithfulness": 0.8813186137735038,
"semantic_formatting": 0.6821703978529401,
"visual_grounding": 0.7045478807517503,
"overall": 0.7278932414784277
},
"denominators": {
"tables": 503,
"charts": 568,
"content_faithfulness": 506,
"semantic_formatting": 456,
"visual_grounding": 500
},
"prediction_sha256": "739f3363883697040f2ab125917ada435824046b47d20be5d4c3930af87a9e83",
"seconds": 111.69487654499244
},
"overlap/single_staged": {
"complete": true,
"pages": 2078,
"fresh_page_evaluations": 2078,
"reused_page_scores": 0,
"official_errors": 0,
"scores": {
"charts": 0.5662119814848688,
"visual_grounding": 0.7045672189887794,
"tables": 0.8039370791018328,
"content_faithfulness": 0.8813104792419619,
"semantic_formatting": 0.682417518427061,
"overall": 0.7276888554489008
},
"independent_scores": {
"tables": 0.8039370791018328,
"charts": 0.5662119814848688,
"content_faithfulness": 0.8813104792419619,
"semantic_formatting": 0.682417518427061,
"visual_grounding": 0.7045672189887794,
"overall": 0.7276888554489008
},
"denominators": {
"tables": 503,
"charts": 568,
"content_faithfulness": 506,
"semantic_formatting": 456,
"visual_grounding": 500
},
"prediction_sha256": "a8b14f3146177bfb7ca9dc07263f0efc66ab5dcb08588f6a84d538f026c1c7b4",
"seconds": 110.24167573684826
},
"overlap/triple_pipelined": {
"complete": true,
"pages": 2078,
"fresh_page_evaluations": 2078,
"reused_page_scores": 0,
"official_errors": 0,
"scores": {
"charts": 0.5786260492246408,
"visual_grounding": 0.7047247783238033,
"tables": 0.8247890476268587,
"content_faithfulness": 0.8807721095566149,
"semantic_formatting": 0.685309450055923,
"overall": 0.7348442869575681
},
"independent_scores": {
"tables": 0.8247890476268587,
"charts": 0.5786260492246408,
"content_faithfulness": 0.8807721095566149,
"semantic_formatting": 0.685309450055923,
"visual_grounding": 0.7047247783238033,
"overall": 0.7348442869575681
},
"denominators": {
"tables": 503,
"charts": 568,
"content_faithfulness": 506,
"semantic_formatting": 456,
"visual_grounding": 500
},
"prediction_sha256": "a216fdafe717e66b495e60f3f9ed92d7dd502f82a3a272d01986151a09ebf23a",
"seconds": 111.52612566621974
},
"overlap/triple_staged": {
"complete": true,
"pages": 2078,
"fresh_page_evaluations": 2078,
"reused_page_scores": 0,
"official_errors": 0,
"scores": {
"charts": 0.5788412927057294,
"visual_grounding": 0.7059309638992248,
"tables": 0.8268812313288487,
"content_faithfulness": 0.8807957411999627,
"semantic_formatting": 0.6831443877791764,
"overall": 0.7351187233825884
},
"independent_scores": {
"tables": 0.8268812313288487,
"charts": 0.5788412927057294,
"content_faithfulness": 0.8807957411999627,
"semantic_formatting": 0.6831443877791764,
"visual_grounding": 0.7059309638992248,
"overall": 0.7351187233825884
},
"denominators": {
"tables": 503,
"charts": 568,
"content_faithfulness": 506,
"semantic_formatting": 456,
"visual_grounding": 500
},
"prediction_sha256": "8f832910cb451b4cb52b62d5efddaa87956d1b93052ec99a71dbcac4002cd436",
"seconds": 112.21417835800094
}
},
"scorer_revision": "9af4102e26b51f234ab5f037e19fec05c0696ec6",
"source_hashes_verified": 241,
"gpu_utilization_snapshots": [
{
"observed_at_utc": "2026-10-17T22:45:26.944779+00:00",
"gpu_utilization_percent": 98,
"power_watts": 680.93,
"memory_used_mib": 78925,
"scope": "one live snapshot; not a run average"
},
{
"observed_at_utc": "2026-10-17T22:59:45.983785+00:00",
"stage": "triple_pipelined",
"gpu_utilization_percent": 100,
"power_watts": 665.11,
"memory_used_mib": 79869,
"scope": "one live snapshot during triple-pass inference; not a run average"
}
],
"artifact_check": {
"files": 77,
"bytes": 49047529,
"per_page_artifacts": false
},
"limitations": [
"Nonzero-temperature generation can differ with batching; request and completion-token totals are reported.",
"Token-limited responses are retained only through explicit allow_partial=True; affected pages are counted incomplete.",
"Prepared throughput includes crop preparation and rendering; excludes startup, initial preparation, warmup and serialization.",
"These measurements are serving profiles, not a hardware throughput ceiling."
],
"merge_only_replays_on_baseline_triple_responses": {
"keep_original_charts": {
"scores": {
"charts": 0.575614863422962,
"visual_grounding": 0.7048420250334769,
"tables": 0.8234070433681294,
"content_faithfulness": 0.8808105394777015,
"semantic_formatting": 0.6812589701148343,
"overall": 0.7331866882834208
},
"fresh_page_evaluations": 2078,
"official_errors": 0,
"replay": {
"pages": 2078,
"variant": "keep_original_charts",
"seconds": 3.7608036780147813,
"overlapping_alternative_blocks_inserted": 0,
"new_inference_requests": 0
}
},
"reference_merge": {
"scores": {
"charts": 0.6462693458732192,
"visual_grounding": 0.7065754139322148,
"tables": 0.8442506317977139,
"content_faithfulness": 0.8809734768808889,
"semantic_formatting": 0.6813144713860899,
"overall": 0.7518766679740253
},
"fresh_page_evaluations": 2078,
"official_errors": 0,
"replay": {
"pages": 2078,
"variant": "reference_merge",
"seconds": 5.24810417900153,
"overlapping_alternative_blocks_inserted": 1664,
"new_inference_requests": 0
}
},
"reference_merge_complete_crops": {
"scores": {
"charts": 0.6466214585492754,
"visual_grounding": 0.7065754139322148,
"tables": 0.8442506317977139,
"content_faithfulness": 0.8809734768808889,
"semantic_formatting": 0.6813144713860899,
"overall": 0.7519470905092366
},
"fresh_page_evaluations": 2078,
"official_errors": 0,
"replay": {
"pages": 2078,
"variant": "reference_merge_complete_crops",
"seconds": 6.634797670994885,
"overlapping_alternative_blocks_inserted": 1660,
"new_inference_requests": 0
}
}
},
"quality_interpretation": "Reference merging can preserve overlapping alternative tables/charts. The chart evaluator searches all rendered tables for the best matching answer. General document output uses guarded replacement and rejects overlapping same-class additions; these are different output contracts.",
"warmup_and_selection": {
"timed_full_runs_start_after": [
"8-page canary for each mode",
"256-page double/triple concurrency pilots"
],
"canary_pages_per_mode": 8,
"pilot_pages": 256,
"pilot_limits": [
128,
256
]
}
}