Datasets:

Vintage-LLM
/

Punch

doc_id large_stringlengths 16 37	item large_stringclasses 156 values	kind large_stringclasses 2 values	volume int64 1 144	year int64 1.84k 1.91k	page int64 0 1.13k	leaf large_stringlengths 16 37	text large_stringlengths 1 22.5k	text_prose large_stringlengths 0 16.2k ⌀	non_prose_spans listlengths 0 97	n_words int64 1 3.78k	prose_words int64 0 3.78k	witness_support float64 0 1 ⌀	dup_of large_stringlengths 16 36 ⌀	flags listlengths 0 4	train bool 2 classes
sim_punch_1841-1846_1_0000	sim_punch_1841-1846_1	sim_microfilm	1	1,841	0	sim_punch_1841-1846_1_0000	PUNCH AND LONDON CHARIVARI VOLUME THE FIRST. LONDON: PUBLISHED FOR THE PROPRIETORS, AT THE OFFICE, 85, FLEET STREET, AND SOLD BY ALL BOOKSELLERS.	LONDON: PUBLISHED FOR THE PROPRIETORS, AT THE OFFICE, 85, FLEET STREET, AND SOLD BY ALL BOOKSELLERS.	[ { "start": 0, "end": 26, "kind": "list" }, { "start": 28, "end": 45, "kind": "list" } ]	23	16	null	null	[]	false
sim_punch_1841-1846_1_0001	sim_punch_1841-1846_1	sim_microfilm	1	1,841	1	sim_punch_1841-1846_1_0001	LONDON BRADURY AND EVANS, PRINTERS, WHITEFRIARS.	null	[]	6	6	null	null	[]	false
sim_punch_1841-1846_1_0002	sim_punch_1841-1846_1	sim_microfilm	1	1,841	2	sim_punch_1841-1846_1_0002	INTRODUCTION. HIS GUFFAWGRAPH is intended to form a refuge for destitute wit—an asylum for the thousands of orphan jokes—the superannuated Joe Millers—the millions of perishing puns, which are now wandering about without so much as a shelf to rest upon! It is also devoted to the emancipation of the Jew d'esprits all o...	null	[]	203	203	0.950495	null	[]	true
sim_punch_1841-1846_1_0003	sim_punch_1841-1846_1	sim_microfilm	1	1,841	3	sim_punch_1841-1846_1_0003	PUNCH'S INTRODUCTION REVIEWS. To render this branch of the periodical as perfect as possible, arrangements have been made to secure the critical assistance of John Ketch, Esq., who, from the mildness of the law, and the congenial character of modern literature with his early associations, has been induced to undertak...	null	[]	210	210	0.935484	null	[]	true
sim_punch_1841-1846_1_0004	sim_punch_1841-1846_1	sim_microfilm	1	1,841	4	sim_punch_1841-1846_1_0004	FOR THE WEEK ENDING JULY 17, 1841. THE MORAL OF PUNCH. As we hope, gentle public, to pass many happy hours in your society, we think it right that you should know something of our character and intentions. Our title, at a first glance, may have misled you into a belief that we have no other intention than the amuseme...	null	[]	1,660	1,660	0.967624	punchvol100a101lemouoft_0397	[]	false
sim_punch_1841-1846_1_0005	sim_punch_1841-1846_1	sim_microfilm	1	1,841	5	sim_punch_1841-1846_1_0005	A RAILROAD NOVEL. Dear Punch,—I was much amused the other day, on taking my seat in the Birmingham Railway train, to observe a sentimental-looking young gentleman, who was sitting opposite to me, deliberately draw from his travelling-bag three volumes of what appeared to me a new novel of the full regulation size, and...	null	[]	996	996	0.982938	null	[]	true
sim_punch_1841-1846_1_0006	sim_punch_1841-1846_1	sim_microfilm	1	1,841	6	sim_punch_1841-1846_1_0006	CONSIDERING TO THE CATEGORIES OF "CANT." It is necessary, therefore, to the full development of the art, that it be brought into such an exposition, as to be able to arrange his polling books according to the different categories, and the bribed to see in what class he shall most properly express it as to bring it (as...	null	[]	1,509	1,509	0.024911	null	[ "reread" ]	true
sim_punch_1841-1846_1_0007	sim_punch_1841-1846_1	sim_microfilm	1	1,841	7	sim_punch_1841-1846_1_0007	THE PROFESSIONAL SINGER Have any of PUNCH's readers ever met one of the above genus—or rather, have they not? They must; for the race is imbued with the most persevering hic et ubique powers. Like the old' mole, these Truepennies "work i' th' dark:" at the Theatres, the Opera, the Coal Hole, the Cider Cellars, and the...	null	[]	1,901	1,901	0.972064	null	[]	true
sim_punch_1841-1846_1_0008	sim_punch_1841-1846_1	sim_microfilm	1	1,841	8	sim_punch_1841-1846_1_0008	A CONVERSATION BETWEEN TWO HACKNEY-COACH HORSES. KINDLY COMMUNICATED BY OUR DOG "TOBY." DEAR SIR,—I was a-sitting the other evening at the door of my kennel, thinking of the dog-days and smoking my pipe (blessings on you, master, for teaching me that art!), when one of your prospectuses was put into my paw by a spanie...	A CONVERSATION BETWEEN TWO HACKNEY-COACH HORSES. KINDLY COMMUNICATED BY OUR DOG "TOBY." DEAR SIR,—I was a-sitting the other evening at the door of my kennel, thinking of the dog-days and smoking my pipe (blessings on you, master, for teaching me that art!), when one of your prospectuses was put into my paw by a spanie...	[ { "start": 8533, "end": 8608, "kind": "list" } ]	1,556	1,544	0.934052	null	[]	true
sim_punch_1841-1846_1_0009	sim_punch_1841-1846_1	sim_microfilm	1	1,841	9	sim_punch_1841-1846_1_0009	STREET POLITICS. A DRAMATIC DIALOGUE BETWEEN PUNCH AND HIS STAGE MANAGER. (Enter Punch.) Punch.—R-r-r-roo-to-toot-tootit? (Sings.) "Wheel about and turn about, And do jes so; Ebery time I turn about, I jump Jim Crow." Manager.—Hollo, Mr. Punch! your voice is rather husky today. Punch.—Yes, yes; I've been making mysel...	null	[]	1,237	1,237	0.933759	null	[]	true
sim_punch_1841-1846_1_0010	sim_punch_1841-1846_1	sim_microfilm	1	1,841	10	sim_punch_1841-1846_1_0010	CANDIDATES UNDER DIFFERENT PHASES. CANVASSING. What a love of a child THE DEPUTATION. Why you want me if you think me THE SUCCESSFUL CANDIDATE. Constituents rascals. THE HUSTINGS. Don't mention it I beg. THE PUBLIC DINNER. The proudest moment of my life.	null	[]	42	42	null	null	[]	true
sim_punch_1841-1846_1_0011	sim_punch_1841-1846_1	sim_microfilm	1	1,841	11	sim_punch_1841-1846_1_0011	CANDIDATES UNDER DIFFERENT NAMES	null	[]	4	4	null	null	[]	false
sim_punch_1841-1846_1_0012	sim_punch_1841-1846_1	sim_microfilm	1	1,841	12	sim_punch_1841-1846_1_0012	FINE ARTS. PUNCH begs most solemnly to assure his friends and the artists in general, that should the violent cold with which he has been from time immemorial afflicted, and which, although it has caused his voice to appear like an infant Lablache screaming through horse-hair and thistles, yet has not very materially ...	FINE ARTS. PUNCH begs most solemnly to assure his friends and the artists in general, that should the violent cold with which he has been from time immemorial afflicted, and which, although it has caused his voice to appear like an infant Lablache screaming through horse-hair and thistles, yet has not very materially ...	[ { "start": 3426, "end": 3457, "kind": "list" }, { "start": 3459, "end": 3898, "kind": "list" } ]	1,347	1,270	0.966165	null	[]	true
sim_punch_1841-1846_1_0013	sim_punch_1841-1846_1	sim_microfilm	1	1,841	13	sim_punch_1841-1846_1_0013	ON THE INTRODUCTION OF PANTOMIME INTO THE ENGLISH LANGUAGE. ENGLISH—it has been remarked a thousand and odd times—is one of the few languages which is unaccompanied with gesticulation. Your veritable Englishman, in his discourse, is as chary as your genuine Frenchman is prodigal, of action. The one speaks like an orac...	null	[]	1,550	1,550	0.976882	null	[]	true
sim_punch_1841-1846_1_0014	sim_punch_1841-1846_1	sim_microfilm	1	1,841	14	sim_punch_1841-1846_1_0014	AN ODE. PICKED UP IN KILLPACK'S DIVAN. Cum notis variorum. "EXCISE COURT.—An information was laid against Mr. Killpack, for selling spirituous liquor. Mr. James (the counsel for the defendant) stated that there was a club held there, of which Mr. Keeley, the actor, was treasurer, and many others of the theatrical prof...	null	[]	1,050	1,050	0.960861	null	[]	true
sim_punch_1841-1846_1_0015	sim_punch_1841-1846_1	sim_microfilm	1	1,841	15	sim_punch_1841-1846_1_0015	THE DRAMA. The "star system" has added another victim to the many already sacrificed to its rapacity and injustice. Mr. Phelps, an actor whose personation of Macduff, the Hunchback, Jaques, &c., would have procured for him in former times no mean position, has been compelled to secede from the Haymarket Theatre from a...	null	[]	1,505	1,505	0.977166	null	[]	true
sim_punch_1841-1846_1_0016	sim_punch_1841-1846_1	sim_microfilm	1	1,841	16	sim_punch_1841-1846_1_0016	FOR THE WEEK ENDING JULY 24, 1841. A MODEST METHOD OF FORMING A NEW BUDGET. SO AS TO PROVIDE FOR THE DEFICIENCY OF THE REVENUE. OOR Mr. Dyer! And so this gentleman has been dismissed from the commission of the peace for humanely endeavouring to obtain the release of Medhurst from confinement. Two or three thousand po...	null	[]	1,224	1,224	0.973222	null	[]	true
sim_punch_1841-1846_1_0017	sim_punch_1841-1846_1	sim_microfilm	1	1,841	17	sim_punch_1841-1846_1_0017	SPORTING. THE KNOCKER HUNT. On Thursday, July 8, 1841, the celebrated pack of Knocker Boys met at the Cavendish, in Jermyn Street. These animals, which have acquired for themselves a celebrity as undying as that of Tom and Jerry, are of a fine powerful breed, and in excellent condition. The success which invariably a...	null	[]	1,177	1,177	0.97649	null	[]	true
sim_punch_1841-1846_1_0018	sim_punch_1841-1846_1	sim_microfilm	1	1,841	18	sim_punch_1841-1846_1_0018	STENOTYPGRAPHY. Our printer's devil, with a laudable anxiety for our success, has communicated the following pathetic story. As a specimen of stenotypography, or compositor's short-hand, we consider it unique. SERAPHINA POPPS; OR, THE BEAUTY OF BLOOMSBURY. Seraphina Pops was the daughter of Mr. Hezekiah Pops, a high...	STENOTYPGRAPHY. Our printer's devil, with a laudable anxiety for our success, has communicated the following pathetic story. As a specimen of stenotypography, or compositor's short-hand, we consider it unique. SERAPHINA POPPS; OR, THE BEAUTY OF BLOOMSBURY. Seraphina Pops was the daughter of Mr. Hezekiah Pops, a high...	[ { "start": 2618, "end": 2856, "kind": "list" }, { "start": 2858, "end": 2889, "kind": "list" } ]	1,042	1,000	0.960765	null	[]	true
sim_punch_1841-1846_1_0019	sim_punch_1841-1846_1	sim_microfilm	1	1,841	19	sim_punch_1841-1846_1_0019	A CHAPTER ON BOOTS. "Boots? Boots!" Yes, Boots! we can write upon boots—we can moralise upon boots; we can convert them, as Jacques does the weeping stag in "As You Like It," (or, whether you like it or not,) into a thousand similes. First, for—but, "our sole's in arms and eager for the fray," and so we will at once h...	null	[]	1,789	1,789	0.942592	null	[]	true
sim_punch_1841-1846_1_0020	sim_punch_1841-1846_1	sim_microfilm	1	1,841	20	sim_punch_1841-1846_1_0020	and holding in the noble brute he strides, as a saftey-valve controls the foaming steam; only loosing him at his very pleasure. Look, look! there's the grey filly, with the other made-to-measure feather on her back; do you notice how she has crawled up to the chesnut? Mark, mark! his arms appear to be India-rubber! Me...	null	[]	1,934	1,934	0.951681	null	[]	true
sim_punch_1841-1846_1_0021	sim_punch_1841-1846_1	sim_microfilm	1	1,841	21	sim_punch_1841-1846_1_0021	PUNCH AND PEEL. Arcades ambo. READER.—God bless us, Mr. PUNCH! who is that tall, fair-haired, somewhat parrot-faced gentleman, smiling like a schoolboy over a mess of treacle, and now kissing the tips of his five fingers as gingerly as if he were doomed to kiss a nettle? PUNCH.—That, Mr. Reader, is the great cotton-p...	READER.—God bless us, Mr. PUNCH! who is that tall, fair-haired, somewhat parrot-faced gentleman, smiling like a schoolboy over a mess of treacle, and now kissing the tips of his five fingers as gingerly as if he were doomed to kiss a nettle? PUNCH.—That, Mr. Reader, is the great cotton-plant, Sir Robert Peel; and at t...	[ { "start": 0, "end": 29, "kind": "list" } ]	1,510	1,505	0.951582	null	[]	true
sim_punch_1841-1846_1_0022	sim_punch_1841-1846_1	sim_microfilm	1	1,841	22	sim_punch_1841-1846_1_0022	PUNCH'S PENCILLINGS.—No. II. HERCULES TEARING THESEUS FROM THE ROCK TO WHICH HE HAD GROWN. (MODERNIZED.) APOLLODORUS relates that THESEUS sat so long on a rock, that at length he grew to it, so that when HERCULES tore him forcibly away, he left all the nether part of the man behind him.	null	[]	51	51	0.843137	null	[]	true
sim_punch_1841-1846_1_0023	sim_punch_1841-1846_1	sim_microfilm	1	1,841	23	sim_punch_1841-1846_1_0023	[?]		[ { "start": 0, "end": 3, "kind": "stamp" } ]	1	0	null	null	[ "no_text" ]	false

End of preview. Expand in Data Studio

Punch, or the London Charivari, 1841–1913

A page-level machine-reading archive of selected Internet Archive scans, with a conservatively reviewed training selection. 52,248 rows; 41,727 selected, byte-distinct training strings containing 36,103,740 whitespace-counted words. All original readings and page provenance are retained, including excluded/uncertain material.

Measure	Count
Archive rows / unique IDs	52,248
Full-reading whitespace words	38,260,995
Retained-view whitespace words	37,315,243
Selected training rows / byte-distinct strings	41,727
Training whitespace words	36,103,740
Internet Archive items	156
Parquet shards / columns	8 / 16

Files and loading

The package contains eight decade Parquet shards in data/, this README, STATS.md, manifest.json, and verify.json. No model weights, source images/ZIPs, historical model completions, classifiers or feature index are bundled. File digests in the manifest cover entire files; the manifest excludes its own checksum to avoid self-reference.

The loader split is archive, containing every row. Select Boolean train=true and train on coalesce(text_prose,text). An empty prose string is not a reason to fall back to the original reading.

From the dataset root, with DuckDB:

SELECT doc_id, coalesce(text_prose, text) AS training_text
FROM read_parquet('data/*.parquet')
WHERE train = true;

With Hugging Face Datasets:

from datasets import load_dataset
archive = load_dataset(
    "parquet", data_files={"archive": "data/*.parquet"}
)["archive"]
selected = archive.filter(lambda r: r["train"])
selected = selected.map(lambda r: {
    "training_text": r["text"] if r["text_prose"] is None else r["text_prose"]
})

Source and reading provenance

There are 54 library items / 35,205 image rows and 102 Serials in Microfilm items / 17,043 image rows. Library contributors include University of Toronto/Robarts, Lincoln, BYU, UIUC, University of Michigan, Google and Servants of Knowledge. Some scans bind multiple volumes and some overlap. Stored year spans 1841–1913; volume/year are inherited selection labels, not a claim that every individual issue date was newly verified or that every printed historical page survives.

The inherited reading lineage is allenai/olmOCR-2-7B-1025-FP8, historically described as vLLM, temperature 0, greyscale page images at 2,200 px longest side and a 5,000-token output cap. Cap/omission candidates were reportedly read as halves or quarters and compared with the IA OCR witness. No model was rerun for this package. Original completions, output-token logs, candidate witness matches and the exact historical programs are unavailable here; these production details are lineage, not fresh verification claims. Witness text was not used as replacement OCR in the conservative repairs.

IA metadata identifies ABBYY or ABBYY-conversion lineage for 48 library items, Tesseract for six library items, and Tesseract/400 ppi for all 102 microfilm items. Engine metadata is not a fresh corpus-wide witness-text download. The stored witness_support values are inherited match fractions for the raw reading, not scores of the revised prose, probabilities of correctness or a substitute for checking the image. Null support does not mean the reading is necessarily wrong.

Retrieve the correct image

Use item and exact leaf, not a guess from the doc_id or ordinal page. All selected source image members follow:

ZIP:    <item>_jp2.zip
member: <item>_jp2/<leaf>.jp2

The source ZIP and witness can be found at:

https://archive.org/download/<item>/<item>_jp2.zip
https://archive.org/download/<item>/<item>_djvu.xml
https://archive.org/metadata/<item>

For an individual BookReader image, fetch item metadata and use a metadata server (try d2, then d1) with BookReader/BookReaderImages.php and parameters zip=<dir>/<item>_jp2.zip, file=<item>_jp2/<leaf>.jp2, id=<item>, scale=1, rotate=0. A generic page/nN URL can select a different leaf, especially for library scans.

Independent enumeration of all 156 primary ZIP central directories found exactly 52,248 image-member names, matching every (item,leaf) once. This verifies the selected member inventory, not every image byte/CRC or OCR transcription.

Fields

field	meaning
`doc_id`	Stable item-plus-zero-based-page ordinal identifier. Not universally identical to leaf.
`item`	Internet Archive item identifier.
`kind`	Scan category: library or sim_microfilm.
`volume`	Inherited selected-volume label; a bound scan may contain additional volumes. Not a new page-level volume adjudication.
`year`	Inherited selection/volume year, not necessarily the printed issue date on every image.
`page`	Contiguous zero-based ordinal within this selected item.
`leaf`	Exact image-member basename without .jp2; authoritative source-image reference.
`text`	Original retained model reading, unchanged; includes OCR errors and archived unreliable generations.
`text_prose`	Derived retained view; null when no spans, empty string when fully removed.
`non_prose_spans`	Removal annotations with Unicode-character [start,end) offsets in text. Use their union; overlaps can occur.
`n_words`	len(text.split()): whitespace tokens, not model tokens or ASCII-regex words.
`prose_words`	Whitespace tokens in coalesce(text_prose,text), including archived non-training readings.
`witness_support`	Nullable inherited raw-reading/IA-OCR token-match fraction. Not a probability, fresh image check or score of revised prose.
`dup_of`	Nullable inherited near-duplicate representative doc_id. Near-cluster assignments unchanged; exact training exclusions use a flag instead.
`flags`	Inherited annotations plus explicit conservative-review/exact-dedup flags; see README definitions.
`train`	Boolean selection under the documented rule. Loader archive split is not this subset.

The 16-column Arrow schema, metadata, original text, original word counts, source identifiers/order, selection labels, witness scores and inherited near-duplicate targets are unchanged. Only five derived fields can differ: non_prose_spans, text_prose, prose_words, flags, train.

Retained view and span semantics

Offsets are sorted Unicode-character [start,end) intervals in the original text, not UTF-8 byte offsets. Remove the union of the intervals, collapse three or more consecutive newlines to two, then strip surrounding whitespace. If there are no intervals, text_prose is null. If everything is removed it is ''. “Prose” is a derived view, not certification that every retained character is authentic printed prose; poetry, dialogue and captions can remain.

head covers inherited running heads/folios, stamp scanner marks and uncertain-only paragraphs, table table markup, and list heuristic lists/advert blocks/calendars/indexes. repeat marks the one source-confirmed repeated caption block cut without modifying the original reading. Historical prose heuristics were not rerun; only adjudicated cuts were reversed.

span kind	spans	annotated chars	span-local whitespace words
head	7,611	333,112	51,459
list	24,841	5,378,063	853,993
repeat	1	342	66
stamp	4,305	21,497	5,458
table	347	357,206	37,137

Kind totals count each annotation separately. Nested table/list spans occur on 18 pages and double-count 25,243 characters across kinds. Union removal covers 6,064,977 original-text characters, whereas annotated lengths sum to 6,090,220. Word sums within spans are also annotation-local and need not equal the difference between full and retained-view word totals. Newline collapse/stripping additionally affects view character totals.

Training selection and flags

train = prose_words >= 20
        and dup_of is null
        and no flag in {still_capped, non_latin_script,
                        review_unanchored, review_fragmentary,
                        review_repetition, review_image_duplicate,
                        exact_train_duplicate}

All word counts are Python whitespace token counts (len(s.split())), not model tokens, ASCII-regex words or a dictionary-based word count. Isolated letters/punctuation can count as tokens, so the word floor alone is not a prose-quality guarantee. There is no blanket 100-character floor, rejection of null support/low support, advertisement ban, or Greek-script ban.

flag	pages	meaning
exact_train_duplicate	29	Literal final training-string copy of another selected row; excluded, archive retained.
loop_cut	139	Inherited cap-triggered suffix cutting, informational; not proof all repetition was removed.
low_support	418	Inherited low witness match, informational; not an automatic exclusion.
no_text	5,067	Zero retained-view words; can result from an empty reading or an entirely removed view.
non_latin_script	9	Inherited Cyrillic/Arabic/CJK screen; excluded. Does not mean every non-Latin script.
reread	2,217	Inherited: a replacement whole/half/quarter model reading was selected; not a fidelity certificate.
review_fragmentary	9	Nine source-checked damaged short letter/word-fragment readings; excluded.
review_image_duplicate	1	One source-confirmed repeat of a printed cartoon/date with an OCR spacing variant; excluded, both source rows retained.
review_repetition	1	Source-confirmed repeated handwritten-caption block; exact copy cut, but remaining reading uncertified and excluded.
review_unanchored	7	Seven source-confirmed extended readings from blank/bleed-through/neighbouring-page material; excluded.
still_capped	123	Inherited retained reading reached the model output cap; excluded.

Counts overlap. There are 7,556 rows with at least one flag and 1,164 inherited near-duplicate exclusions. Near-cluster targets are unchanged. Exact deduplication is a separate training-only operation: hash the final UTF-8 view, confirm literal equality in matching groups, then prefer a library scan, higher non-null inherited support, fewer flags, and lexicographically earlier doc_id. No case-folding, Unicode normalization or punctuation normalization is used. Excluded copies remain in the archive; every exact_train_duplicate has a selected literal-equal counterpart.

Conservative source-backed review

selection/annotation action	rows	word effect
Unanchored-generation exclusion	7	−1,986 training words
Fragmentary-reading exclusion	9	−254 training words
Repeat cut and precautionary exclusion	1	−66 retained-view words; −138 training words
Source-checked narrative restoration	6	+207 retained-view and training words (8 spans)
Source-confirmed image-copy exclusion	1	−22 training words
Exact final-string exclusion	29	−942 training words

Only 53 archive rows have changed derived values: flags/selection change on 47 rows and spans/view/counts on seven (one overlaps). Net retained-view gain is 141 words; training changes by −47 rows / −3,135 words. These are repair effects, not a corpus-wide error estimate.

Reviewed a targeted pool of 43 long null-support training readings: seven confirmed unanchored generations were excluded, while the other 36 were left selected. Many are genuine poetry, captions, almanack text or multi-column articles. This was a source-presence/layout review, not a word-error-rate benchmark.
Checked all 16 selected readings shorter than 101 characters: nine fragmentary near-blank readings were excluded; seven short caption readings passed this content check. One was subsequently excluded as a source-confirmed image copy, leaving six selected. No general length-based deletion was applied.
punchvol98a99lemouoft_0559 has the exact repeated block at [434,776) removed (66 words), leaving 72 view words. It remains excluded because the handwritten-caption transcription was not certified after cutting.
Restored eight false-positive list spans on six pages, retaining the model characters verbatim, including two associated article headings:

doc_id	restored intervals	restored words
`punchvol140a141lemouoft_0574`	[0,24), [26,313)	51
`punch1381unse_0639`	[3129,3155), [3157,3371)	39
`sim_punch_july-december-1912_143_0347`	[6125,6304)	25
`sim_punch_july-december-1912_143_0349`	[0,217)	34
`punchvol144a145lemouoft_0087`	[4455,4669)	33
`punchvol124a125lemouoft_0254`	[48,206)	25

The one apparent near-cluster recovery candidate was not admitted: its extra 37-word view is masthead/advertising residue, not source-validated narrative.
Exact images for punch7018unse_0271 and punch70a71lemouoft_0264 show the same THE DOGS OF WAR cartoon dated June 17, 1876. Their captions differ in BULL A1 versus BULL A 1, escaping literal-string deduplication. The latter 22-word copy is withheld with review_image_duplicate; both archive rows and their inherited near-cluster targets remain unchanged.
Removed 29 redundant exact training copies / 942 words from selection only.

Source checks address 69 distinct exact-leaf images including the earlier audit controls. They are targeted checks, not a random sample, full transcription adjudication or validation of every cut. The 87-candidate readable-cut screen was a review aid; unaccepted candidates remain cut. No speculative replacement OCR was written, and all inherited flags were preserved.

Current counts

decade	pages	words	prose_words	train_pages	train_words
1840s	4,734	4,424,493	4,350,629	3,832	4,305,207
1850s	5,653	5,088,006	4,984,129	4,508	4,933,202
1860s	6,423	5,416,882	5,291,381	4,726	4,761,147
1870s	6,380	4,626,556	4,537,296	5,032	4,321,803
1880s	7,477	5,570,588	5,320,804	6,130	5,230,185
1890s	7,319	4,850,878	4,701,015	5,612	4,445,839
1900s	9,942	5,805,465	5,691,737	8,240	5,676,267
1910s	4,320	2,478,127	2,438,252	3,647	2,430,090

scan	items	pages	words	prose_words	train_pages	train_words
library	54	35,205	24,454,133	23,952,572	27,717	22,920,414
sim_microfilm	102	17,043	13,806,862	13,362,671	14,010	13,183,326

The complete 156-item and stored-year tables, fresh training-string statistics and raw-support distributions are in STATS.md, with machine-readable counterparts in the manifest. The current 41,728 selected rows are 41,728 byte-distinct final strings. No unavailable historical external-index/dictionary/entropy/quality feature table is presented as a fresh census.

Verification and limits

verify.json records independent read-back checks of every written row: identical schema/metadata, row order/IDs, all 11 fixed fields, only approved derived changes, full/prose token counts, every span-union reconstruction, flags/eligibility, near-duplicate targets, source-member inventory, exact final-string uniqueness, and literal-equal counterparts for exact exclusions. Package checks compare all tables/totals against the Parquets and rehash all inventoried files.

Zero failures in the recorded fresh checks. This means structural/selection/documentation consistency, not perfect OCR. The model can invent fluent passages from blank margins, bleed-through or slivers of facing pages, omit columns and misread names/dates/handwriting; some such defects can remain outside the bounded review. Witness matches and English lexical plausibility do not rule out hallucinations. A reread can also repeat content. Source images, all raw model completions, historical classifiers, fingerprints and feature-index artifacts are not bundled, and their old generative/recall tests were not newly reproduced. Printed dates can differ from inherited volume-year labels, especially almanacks. Exact-string deduplication does not guarantee elimination of every OCR-variant duplicate or historical reprint.

Licence and credits

The transcription/annotation dataset carries the inherited CC0 1.0 dedication. Source scans and underlying printed contributions have their own rights context; IA copyright evidence commonly specifies US. Publication before 1914 does not establish worldwide public-domain status under every jurisdiction's author-life/local term rules. CC0 cannot waive rights held by others; this package does not certify worldwide rights clearance. Credit Internet Archive, Serials in Microfilm and the library contributors listed above. The package was prepared locally, without publication/upload or paid/generative OCR calls.

Downloads last month: 21