validation_report.md

Validation report — data v1

ALL CHECKS PASS

• generated by python -m data.validate --config configs/data/v1.yaml
• config fingerprints: pile 3f4303d9b926b730, bios 7a8c48a96792d725, pack 15643c90477c31a7
• Pile revision 3be90335b66f24456a5d6659d9c8d208c0357119
• bundle: 59 files, 1,151.6 MB (753.6 MB without raw/)
• git ff1bc917cb40

check	what	result	measured
A1	Rebuild is byte-identical	PASS	59 files hashed; 0 changed, 0 added, 0 removed
A2	data.fetch --verify-only passes	PASS	59/59 files verified against v1.json (0 mismatched, 0 missing)
A3	Token counts per subject and half	PASS	every Pile subject exactly 27,000,000 in train.bin; every held-out half within 1% of 2,500,000 per subject
A4	Shares in every 1,000,000-token window	PASS	worst deviation 0.416 points across 598 overlapping windows (stride 250,000), tolerance 2.0
A5	Format: dtype, vocab bound, EOT, index tiling	PASS	train.bin: 150,475,264 tokens, 332,886 segments, 333,027 EOT separator/pad tokens, max id 50256 | label.bin: 15,021,806 tokens, 13,306 segments, 13,305 EOT separator/pad tokens, max id 50256 | test.bin: 15,023,952 tokens, 13,381 segments, 13,380 EOT separator/pad tokens, max id 50256
A6	Decoding index segments reproduces the source	PASS	1,000 random segments checked token for token (820 of them whole documents, also compared as text); 0 mismatches
A7	Leaks: ids, exact hashes, 13-gram overlap	PASS	ids disjoint across train/label/test; 0 exact-hash matches; 0 documents over 50%; worst single document 50.0%; aggregates Pile-CC=0.64%, Github=14.35%, PubMed=2.93%, FreeLaw=9.21%, Wikipedia=2.14%, USPTO=0.00%
A8	Halves: disjoint, >=500 docs, <=4096 tokens	PASS	halves share 0 documents; smallest half/subject is 708 documents (minimum 500); longest document 4096 tokens (cap 4096)
A9	No held-out-only subject in train.bin	PASS	['USPTO Backgrounds'] appear in 0 train.bin segments by document id and 0 by subject label; train.bin carries ['FreeLaw', 'Github', 'Pile-CC', 'PubMed Central', 'Wikipedia (en)', 'bio']
A10	Subject labels agree with the source	PASS	142,886 Pile segments: index subject == raw-archive meta.pile_set_name == the document id's slug; 190,000 biography segments all labelled bio with a known person id
A11	Bag-of-words subject classifier (sanity)	PASS	accuracy 94.87% on 13,381 test-half documents (threshold 90%); per subject Pile-CC=89.7%, Github=97.8%, PubMed=97.6%, FreeLaw=99.1%, Wikipedia=95.0%, USPTO=96.2%
A12	Biography facts	PASS	2,500 people, names unique and absent from the Pile, genders {'female': 1250, 'male': 1250}, 5 facts each; first-token uniqueness holds per list (first two for birth_date, which has only 12 distinct first tokens); chi-square p birth_date=0.124, birth_city=0.877, university=0.183, major=0.940, employer=0.347
A13	Fact prompts: no answer leaks	PASS	100,000 rows: 0 leak the value verbatim (0), 0 via the person's name (0), 0 via another of the person's values (0); 50,000 training prompts are exact prefixes of their biography; 50,000 paraphrase prompts, 200 of them searched in full against train.bin and absent; held-out markers inside biography text: 'Question:'=0, 'When it comes to'=0, 'is on record as'=0, 'reads as follows'=0 (the same markers occur 571 times in ordinary Pile text, which is not a leak); REPORTED: 48 rows whose ranking key is a TEMPLATE word (birth_city=Thessaloniki, Greece, employer=Heineken, major=Theology, university=Aalborg University)
A14	Exposure counts and control isolation	PASS	190,000 copies over 2,000 trained people, every count equal to its tier target and spread over the 4 rewordings; 0 cross a 512-token boundary; 0 of 500 control names appear in train text
A15	Probe floor: random and untrained sit mid-list	PASS	random scorer: mean rank fraction 0.502, top-1 0.558% over 36,000 prompts; untrained model: 0.501, top-1 0.525% over 4,000; mean candidate list 210
A16	Loader: order, resume, shapes	PASS	293,896 sequences of 512; same seed reproduces the order, a different seed does not, the order is a permutation, and a resume from (seed, position) matches an unbroken run
A17	Model beats the per-subject unigram baseline	PASS	Pile-CC: model 6.7886 vs unigram 7.559; Github: model 5.7371 vs unigram 6.3888; PubMed: model 6.5792 vs unigram 7.4554; FreeLaw: model 5.8735 vs unigram 6.8493; Wikipedia: model 7.0163 vs unigram 7.72; USPTO: model 7.0021 vs unigram 7.9605
A18	Biography facts are learnable	PASS	5,830 steps / 96M tokens in 2.03 h of training. test top-1: frequent 40.5%, normal 3.1%, rare 0.8%, control 0.6% (frequent/control = 64.003x, needs 2x; monotone yes). paraphrase top-1: frequent 2.2%, control 0.4%

Timings

check	seconds
A1	253.0
A2	0.4
A3	0.0
A4	2.9
A5	0.3
A6	2.8
A7	88.3
A8	0.0
A9	0.0
A10	4.9
A11	0.5
A12	0.0
A13	8.0
A14	3.4
A15	10.0
A16	0.1
A17	7.0
A18	0.0

Total 382 s.