{ "timing": { "completed": true, "total_tokens": 7620187, "elapsed_seconds": 94.359459165 }, "before": { "categories": { "ecb": { "before_loss": 2.5478364897389945, "after_loss": 2.494417235082108, "delta_loss": -0.05341925465688655, "delta_ci95": [ -0.06972793260970898, -0.04159707502203487 ], "perplexity_ratio_after_before": 0.9479824830844155, "tokens": 712043, "families": 256, "direction": "better" }, "ema": { "before_loss": 2.474508373573441, "after_loss": 2.3786064237293756, "delta_loss": -0.09590194984406535, "delta_ci95": [ -0.11510346025159647, -0.07976488583414702 ], "perplexity_ratio_after_before": 0.9085530954741962, "tokens": 361602, "families": 112, "direction": "better" }, "original/arxiv": { "before_loss": 1.9497882117902077, "after_loss": 1.9627351147577499, "delta_loss": 0.012946902967542164, "delta_ci95": [ 0.0025936547777961195, 0.020426215534584412 ], "perplexity_ratio_after_before": 1.0130310769877917, "tokens": 955479, "families": 256, "direction": "worse" }, "original/code_python": { "before_loss": 1.5591427360983965, "after_loss": 1.4213988499193573, "delta_loss": -0.1377438861790393, "delta_ci95": [ -0.16061217365644495, -0.11849704440245086 ], "perplexity_ratio_after_before": 0.8713218207321642, "tokens": 346847, "families": 256, "direction": "better" }, "original/doab": { "before_loss": 2.362370064841509, "after_loss": 2.3794003204124636, "delta_loss": 0.017030255570954583, "delta_ci95": [ 0.012913304267402897, 0.02087139434915268 ], "perplexity_ratio_after_before": 1.0171760971032633, "tokens": 207612, "families": 71, "direction": "worse" }, "original/finemath_4plus": { "before_loss": 1.799093179155612, "after_loss": 1.816838842892399, "delta_loss": 0.017745663736787015, "delta_ci95": [ 0.013976993094340947, 0.021202179749114675 ], "perplexity_ratio_after_before": 1.0179040535511663, "tokens": 324946, "families": 256, "direction": "worse" }, "original/fineweb_edu": { "before_loss": 2.675224465840467, "after_loss": 2.7002269013800166, "delta_loss": 0.025002435539549325, "delta_ci95": [ 0.022486436681909267, 0.027483818841139318 ], "perplexity_ratio_after_before": 1.0253176177229963, "tokens": 308280, "families": 256, "direction": "worse" }, "original/gutenberg": { "before_loss": 2.8188761140753904, "after_loss": 2.8476179976099036, "delta_loss": 0.028741883534513044, "delta_ci95": [ 0.02373568672012666, 0.034234945455012794 ], "perplexity_ratio_after_before": 1.0291589173266178, "tokens": 829715, "families": 211, "direction": "worse" }, "original/library_of_congress": { "before_loss": 2.880806793568109, "after_loss": 2.8955215596255326, "delta_loss": 0.01471476605742327, "delta_ci95": [ 0.012569879553039051, 0.0168030816943064 ], "perplexity_ratio_after_before": 1.0148235612042, "tokens": 620382, "families": 162, "direction": "worse" }, "original/libretexts": { "before_loss": 2.2442110126607084, "after_loss": 2.2929445338363856, "delta_loss": 0.048733521175677315, "delta_ci95": [ 0.041682197188669175, 0.05636858044326998 ], "perplexity_ratio_after_before": 1.049940526540951, "tokens": 145786, "families": 84, "direction": "worse" }, "original/pes2o": { "before_loss": 2.312943282481791, "after_loss": 2.330488094417853, "delta_loss": 0.017544811936062, "delta_ci95": [ 0.01588378921244097, 0.019099961748945963 ], "perplexity_ratio_after_before": 1.0176996262195315, "tokens": 905439, "families": 256, "direction": "worse" }, "original/pre_1929_books": { "before_loss": 2.8731941185621293, "after_loss": 2.888885086531724, "delta_loss": 0.01569096796959501, "delta_ci95": [ 0.013513525804337942, 0.017965919043504367 ], "perplexity_ratio_after_before": 1.0158147176108423, "tokens": 817106, "families": 208, "direction": "worse" }, "original/pressbooks": { "before_loss": 2.5052328276563767, "after_loss": 2.546299820711243, "delta_loss": 0.04106699305486627, "delta_ci95": [ 0.03410804841715809, 0.04900897419272933 ], "perplexity_ratio_after_before": 1.0419219047387533, "tokens": 256223, "families": 142, "direction": "worse" }, "original/pubmed": { "before_loss": 2.1153214027610927, "after_loss": 2.140136179856328, "delta_loss": 0.024814777095235434, "delta_ci95": [ 0.022839311901323153, 0.026803260436568794 ], "perplexity_ratio_after_before": 1.025125226266432, "tokens": 828727, "families": 256, "direction": "worse" } }, "sample": { "identity": { "manifest_sha256": [ "7ecc680b9ca742ca67eae14c072dd0f7a28a3a0a24f885ec7dfd022ffc180778", "50fafc41b4cce78b31922e4a9e37d1bdaa1bf49f009e9e9b4ec1c812a4a8461c" ], "tokenizer_sha256": "add5b2c7542e4efded329735825ecdcb94df1432617c3d3e4b7508aa459432f9", "seed": 20261019, "context": 2048, "parents_per_source": 256, "windows_per_document": 2, "implementation_sha256": "dc8a9a7342f479ae52b2f8d95e80e1ba59a46646a1acb3e9dbafbd13eb40c585" }, "windows_sha256": "579fe4fb51b9165328c1e0be0959307245555df08d8b63f285c81456951bb1db", "coverage": { "original/fineweb_edu": { "available_documents": 24542, "available_families": 24542, "selected_documents": 256, "selected_families": 256, "missing_parent_documents": 0, "windows": 294, "tokens": 308280, "nonprefix_windows": 40 }, "original/code_python": { "available_documents": 38635, "available_families": 10395, "selected_documents": 256, "selected_families": 256, "missing_parent_documents": 0, "windows": 320, "tokens": 346847, "nonprefix_windows": 77 }, "original/finemath_4plus": { "available_documents": 39488, "available_families": 39347, "selected_documents": 256, "selected_families": 256, "missing_parent_documents": 0, "windows": 296, "tokens": 324946, "nonprefix_windows": 50 }, "original/pes2o": { "available_documents": 3966, "available_families": 3966, "selected_documents": 256, "selected_families": 256, "missing_parent_documents": 0, "windows": 508, "tokens": 905439, "nonprefix_windows": 371 }, "original/pubmed": { "available_documents": 1308, "available_families": 1308, "selected_documents": 256, "selected_families": 256, "missing_parent_documents": 0, "windows": 478, "tokens": 828727, "nonprefix_windows": 328 }, "original/arxiv": { "available_documents": 398, "available_families": 398, "selected_documents": 256, "selected_families": 256, "missing_parent_documents": 0, "windows": 508, "tokens": 955479, "nonprefix_windows": 404 }, "original/pressbooks": { "available_documents": 3217, "available_families": 142, "selected_documents": 142, "selected_families": 142, "missing_parent_documents": 0, "windows": 196, "tokens": 256223, "nonprefix_windows": 67 }, "original/libretexts": { "available_documents": 4718, "available_families": 84, "selected_documents": 84, "selected_families": 84, "missing_parent_documents": 0, "windows": 114, "tokens": 145786, "nonprefix_windows": 41 }, "original/doab": { "available_documents": 1045, "available_families": 71, "selected_documents": 71, "selected_families": 71, "missing_parent_documents": 0, "windows": 120, "tokens": 207612, "nonprefix_windows": 91 }, "original/gutenberg": { "available_documents": 211, "available_families": 211, "selected_documents": 211, "selected_families": 211, "missing_parent_documents": 0, "windows": 418, "tokens": 829715, "nonprefix_windows": 394 }, "original/pre_1929_books": { "available_documents": 208, "available_families": 208, "selected_documents": 208, "selected_families": 208, "missing_parent_documents": 0, "windows": 411, "tokens": 817106, "nonprefix_windows": 380 }, "original/library_of_congress": { "available_documents": 162, "available_families": 162, "selected_documents": 162, "selected_families": 162, "missing_parent_documents": 0, "windows": 317, "tokens": 620382, "nonprefix_windows": 283 }, "ecb": { "available_documents": 296, "available_families": 293, "selected_documents": 256, "selected_families": 256, "missing_parent_documents": 0, "windows": 437, "tokens": 712043, "nonprefix_windows": 268 }, "ema": { "available_documents": 244, "available_families": 112, "selected_documents": 112, "selected_families": 112, "missing_parent_documents": 0, "windows": 198, "tokens": 361602, "nonprefix_windows": 140 } }, "total_tokens": 7620187, "total_windows": 4615, "total_families": 2782, "sampling_note": "Uniform document-family sample, one document per family, up to two uniform nonoverlapping 2K blocks per document. Token-weighted loss on this sample is not a whole-corpus token-uniform estimate." }, "before_identity": { "checkpoint_sha256": "ce26b11aed843008fb89893a5e90219b1a5c96e44ef32d42eb9188e427df1876", "sample_sha256": "579fe4fb51b9165328c1e0be0959307245555df08d8b63f285c81456951bb1db", "implementation_sha256": "dc8a9a7342f479ae52b2f8d95e80e1ba59a46646a1acb3e9dbafbd13eb40c585", "precision": "FP32; TF32 disabled" }, "after_identity": { "checkpoint_sha256": "88b039618647c7f0ff7988165c9ae0b0075540927049b2f8e3b625272dfa8dda", "sample_sha256": "579fe4fb51b9165328c1e0be0959307245555df08d8b63f285c81456951bb1db", "implementation_sha256": "dc8a9a7342f479ae52b2f8d95e80e1ba59a46646a1acb3e9dbafbd13eb40c585", "precision": "FP32; TF32 disabled" }, "bootstrap_samples": 2000, "limitations": [ "Validation split only; no test-set tuning or scoring.", "Family-balanced sample, not corpus-token-uniform perplexity.", "CIs are conditional on selected families and are not simultaneous across categories.", "Source split/near-duplicate/contamination audits remain pending.", "2K window loss does not establish 8K long-context quality or generation correctness." ] }, "after": { "categories": { "ecb": { "before_loss": 2.0649489631876783, "after_loss": 2.494417235082108, "delta_loss": 0.42946827189442965, "delta_ci95": [ 0.4070746200382845, 0.45517377462533415 ], "perplexity_ratio_after_before": 1.536440337797081, "tokens": 712043, "families": 256, "direction": "worse" }, "ema": { "before_loss": 1.7661885360415588, "after_loss": 2.3786064237293756, "delta_loss": 0.6124178876878167, "delta_ci95": [ 0.5505569266329057, 0.682683916389956 ], "perplexity_ratio_after_before": 1.8448867392869388, "tokens": 361602, "families": 112, "direction": "worse" }, "original/arxiv": { "before_loss": 2.8091927217655503, "after_loss": 1.9627351147577499, "delta_loss": -0.8464576070078007, "delta_ci95": [ -0.8869131449000369, -0.8087678207312736 ], "perplexity_ratio_after_before": 0.4289316884962148, "tokens": 955479, "families": 256, "direction": "better" }, "original/code_python": { "before_loss": 3.9942717673282786, "after_loss": 1.4213988499193573, "delta_loss": -2.572872917408921, "delta_ci95": [ -2.6817265777019617, -2.475202211644162 ], "perplexity_ratio_after_before": 0.07631598066978175, "tokens": 346847, "families": 256, "direction": "better" }, "original/doab": { "before_loss": 2.837422036187953, "after_loss": 2.3794003204124636, "delta_loss": -0.4580217157754896, "delta_ci95": [ -0.4989884484846478, -0.41974836024816753 ], "perplexity_ratio_after_before": 0.6325337400975293, "tokens": 207612, "families": 71, "direction": "better" }, "original/finemath_4plus": { "before_loss": 2.518498037605622, "after_loss": 1.816838842892399, "delta_loss": -0.7016591947132227, "delta_ci95": [ -0.7774240937365081, -0.6293870040452769 ], "perplexity_ratio_after_before": 0.4957620552343834, "tokens": 324946, "families": 256, "direction": "better" }, "original/fineweb_edu": { "before_loss": 2.9112838428345946, "after_loss": 2.7002269013800166, "delta_loss": -0.2110569414545781, "delta_ci95": [ -0.22441697225249804, -0.19922256652313894 ], "perplexity_ratio_after_before": 0.8097279584807806, "tokens": 308280, "families": 256, "direction": "better" }, "original/gutenberg": { "before_loss": 3.4374800509441092, "after_loss": 2.8476179976099036, "delta_loss": -0.5898620533342056, "delta_ci95": [ -0.6134544118766608, -0.566681093819697 ], "perplexity_ratio_after_before": 0.5544037576096629, "tokens": 829715, "families": 211, "direction": "better" }, "original/library_of_congress": { "before_loss": 3.2692679650982717, "after_loss": 2.8955215596255326, "delta_loss": -0.3737464054727389, "delta_ci95": [ -0.3827009216357588, -0.3650596476897178 ], "perplexity_ratio_after_before": 0.6881514011325324, "tokens": 620382, "families": 162, "direction": "better" }, "original/libretexts": { "before_loss": 2.5554314526110096, "after_loss": 2.2929445338363856, "delta_loss": -0.2624869187746239, "delta_ci95": [ -0.2871515101622813, -0.23976097907801563 ], "perplexity_ratio_after_before": 0.7691364255496916, "tokens": 145786, "families": 84, "direction": "better" }, "original/pes2o": { "before_loss": 2.5647577556418253, "after_loss": 2.330488094417853, "delta_loss": -0.23426966122397241, "delta_ci95": [ -0.24308463848508355, -0.2257324247828331 ], "perplexity_ratio_after_before": 0.7911484450699164, "tokens": 905439, "families": 256, "direction": "better" }, "original/pre_1929_books": { "before_loss": 3.2404382420261166, "after_loss": 2.888885086531724, "delta_loss": -0.35155315549439237, "delta_ci95": [ -0.3608818028641581, -0.3420810440792625 ], "perplexity_ratio_after_before": 0.7035944490571835, "tokens": 817106, "families": 208, "direction": "better" }, "original/pressbooks": { "before_loss": 2.751147606064535, "after_loss": 2.546299820711243, "delta_loss": -0.20484778535329193, "delta_ci95": [ -0.21972552084407188, -0.19178308128206775 ], "perplexity_ratio_after_before": 0.8147713271027561, "tokens": 256223, "families": 142, "direction": "better" }, "original/pubmed": { "before_loss": 2.4947581148259883, "after_loss": 2.140136179856328, "delta_loss": -0.3546219349696603, "delta_ci95": [ -0.3786113121725641, -0.3318440784653726 ], "perplexity_ratio_after_before": 0.7014385824844385, "tokens": 828727, "families": 256, "direction": "better" } }, "sample": { "identity": { "manifest_sha256": [ "7ecc680b9ca742ca67eae14c072dd0f7a28a3a0a24f885ec7dfd022ffc180778", "50fafc41b4cce78b31922e4a9e37d1bdaa1bf49f009e9e9b4ec1c812a4a8461c" ], "tokenizer_sha256": "add5b2c7542e4efded329735825ecdcb94df1432617c3d3e4b7508aa459432f9", "seed": 20261019, "context": 2048, "parents_per_source": 256, "windows_per_document": 2, "implementation_sha256": "dc8a9a7342f479ae52b2f8d95e80e1ba59a46646a1acb3e9dbafbd13eb40c585" }, "windows_sha256": "579fe4fb51b9165328c1e0be0959307245555df08d8b63f285c81456951bb1db", "coverage": { "original/fineweb_edu": { "available_documents": 24542, "available_families": 24542, "selected_documents": 256, "selected_families": 256, "missing_parent_documents": 0, "windows": 294, "tokens": 308280, "nonprefix_windows": 40 }, "original/code_python": { "available_documents": 38635, "available_families": 10395, "selected_documents": 256, "selected_families": 256, "missing_parent_documents": 0, "windows": 320, "tokens": 346847, "nonprefix_windows": 77 }, "original/finemath_4plus": { "available_documents": 39488, "available_families": 39347, "selected_documents": 256, "selected_families": 256, "missing_parent_documents": 0, "windows": 296, "tokens": 324946, "nonprefix_windows": 50 }, "original/pes2o": { "available_documents": 3966, "available_families": 3966, "selected_documents": 256, "selected_families": 256, "missing_parent_documents": 0, "windows": 508, "tokens": 905439, "nonprefix_windows": 371 }, "original/pubmed": { "available_documents": 1308, "available_families": 1308, "selected_documents": 256, "selected_families": 256, "missing_parent_documents": 0, "windows": 478, "tokens": 828727, "nonprefix_windows": 328 }, "original/arxiv": { "available_documents": 398, "available_families": 398, "selected_documents": 256, "selected_families": 256, "missing_parent_documents": 0, "windows": 508, "tokens": 955479, "nonprefix_windows": 404 }, "original/pressbooks": { "available_documents": 3217, "available_families": 142, "selected_documents": 142, "selected_families": 142, "missing_parent_documents": 0, "windows": 196, "tokens": 256223, "nonprefix_windows": 67 }, "original/libretexts": { "available_documents": 4718, "available_families": 84, "selected_documents": 84, "selected_families": 84, "missing_parent_documents": 0, "windows": 114, "tokens": 145786, "nonprefix_windows": 41 }, "original/doab": { "available_documents": 1045, "available_families": 71, "selected_documents": 71, "selected_families": 71, "missing_parent_documents": 0, "windows": 120, "tokens": 207612, "nonprefix_windows": 91 }, "original/gutenberg": { "available_documents": 211, "available_families": 211, "selected_documents": 211, "selected_families": 211, "missing_parent_documents": 0, "windows": 418, "tokens": 829715, "nonprefix_windows": 394 }, "original/pre_1929_books": { "available_documents": 208, "available_families": 208, "selected_documents": 208, "selected_families": 208, "missing_parent_documents": 0, "windows": 411, "tokens": 817106, "nonprefix_windows": 380 }, "original/library_of_congress": { "available_documents": 162, "available_families": 162, "selected_documents": 162, "selected_families": 162, "missing_parent_documents": 0, "windows": 317, "tokens": 620382, "nonprefix_windows": 283 }, "ecb": { "available_documents": 296, "available_families": 293, "selected_documents": 256, "selected_families": 256, "missing_parent_documents": 0, "windows": 437, "tokens": 712043, "nonprefix_windows": 268 }, "ema": { "available_documents": 244, "available_families": 112, "selected_documents": 112, "selected_families": 112, "missing_parent_documents": 0, "windows": 198, "tokens": 361602, "nonprefix_windows": 140 } }, "total_tokens": 7620187, "total_windows": 4615, "total_families": 2782, "sampling_note": "Uniform document-family sample, one document per family, up to two uniform nonoverlapping 2K blocks per document. Token-weighted loss on this sample is not a whole-corpus token-uniform estimate." }, "before_identity": { "checkpoint_sha256": "a9b4ff39352fae4aa163433df7130dbede1437657bee7c44fc26a6eb020af565", "sample_sha256": "579fe4fb51b9165328c1e0be0959307245555df08d8b63f285c81456951bb1db", "implementation_sha256": "dc8a9a7342f479ae52b2f8d95e80e1ba59a46646a1acb3e9dbafbd13eb40c585", "precision": "FP32; TF32 disabled" }, "after_identity": { "checkpoint_sha256": "88b039618647c7f0ff7988165c9ae0b0075540927049b2f8e3b625272dfa8dda", "sample_sha256": "579fe4fb51b9165328c1e0be0959307245555df08d8b63f285c81456951bb1db", "implementation_sha256": "dc8a9a7342f479ae52b2f8d95e80e1ba59a46646a1acb3e9dbafbd13eb40c585", "precision": "FP32; TF32 disabled" }, "bootstrap_samples": 2000, "limitations": [ "Validation split only; no test-set tuning or scoring.", "Family-balanced sample, not corpus-token-uniform perplexity.", "CIs are conditional on selected families and are not simultaneous across categories.", "Source split/near-duplicate/contamination audits remain pending.", "2K window loss does not establish 8K long-context quality or generation correctness." ] } }