Datasets:

biglam
/

londons-pulse-moh

language:
  - en
license: cc-by-nc-4.0
task_categories:
  - image-to-text
  - image-classification
pretty_name: 'London''s Pulse: MOH Reports (page images + OCR text)'
tags:
  - archives
  - glam
  - historical
  - iiif
  - lam
  - libraries
  - medical-officer-of-health
  - ocr
  - public-health
  - table-extraction
  - tables
  - wellcome-collection
dataset_info:
  config_name: tables
  features:
    - name: image
      dtype: image
    - name: page_type
      dtype: string
    - name: table_ground_truth
      dtype: string
    - name: n_tables_on_page
      dtype: int32
    - name: b_number
      dtype: string
    - name: page_index
      dtype: int32
    - name: printed_page
      dtype: int32
    - name: source_collection
      dtype: string
    - name: label_source
      dtype: string
    - name: split
      dtype: string
    - name: signals
      dtype: string
  splits:
    - name: train
      num_bytes: 448527779
      num_examples: 2520
    - name: validation
      num_bytes: 52884161
      num_examples: 302
    - name: test
      num_bytes: 56707410
      num_examples: 327
  download_size: 554397290
  dataset_size: 558119350
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
  - config_name: tables
    data_files:
      - split: train
        path: tables/train-*
      - split: validation
        path: tables/validation-*
      - split: test
        path: tables/test-*

London's Pulse: Medical Officer of Health reports (page images + OCR text)

Page-level scans of the Wellcome Collection London's Pulse Medical Officer of Health (MOH) reports (1848–1972), paired with OCR text, per-report licence, and full provenance. Built for OCR / VLM / document-understanding work on real historical public-health records — dense statistical tables, mixed layouts, century-old print.

Configs

config	rows	what
`default`	391,964 pages / 4,886 reports	every page image + its report's OCR text + licence
`tables`	3,149 pages / 288 reports	selected page images paired with 3,980 machine-extracted tables; silver labels for page-type training and exploratory table-extraction evaluation

from datasets import load_dataset

# full corpus (stream — it's ~110 GB)
ds = load_dataset("biglam/londons-pulse-moh", split="train", streaming=True)

# tables subset (small; image + silver table targets)
tab = load_dataset("biglam/londons-pulse-moh", "tables", split="test")

`default` config — columns

column	type	description
`image`	`Image`	the page scan (full native resolution)
`b_number`	`string`	Wellcome report id (join key)
`page_index`	`int32`	1-based page (scan) number within the report
`n_pages`	`int32`	total pages in the report
`report_text`	`string`	OCR text of the whole report (see caveat)
`license`	`string`	normalised image licence (all `cc-by-nc` here)
`table_ground_truth`	`string`	reserved (empty in `default`; populated in the `tables` config)
`signals`	`string`	JSON provenance (`manifest_url`, `image_service_id`, size/format, `license_raw`, multi-volume linkage)

report_text is report-level, not page-aligned — the source OCR is a flat per-report dump with no reliable page boundaries, so the same full-report text repeats across every page of that report. Group/dedupe by b_number.

`tables` config — columns

One row per scan with at least one matched table from Wellcome's machine-extracted export. The attached tables are a selected subset, not an exhaustive annotation of every table visible on the page. The CSV content remains uncorrected silver data. Use for table page-type training or exploratory OCR/VLM evaluation with that limitation.

column	type	description
`image`	`Image`	the page scan
`page_type`	`string`	`"table"`
`table_ground_truth`	`string`	JSON list of `{table_id, csv}` — the extracted table(s) on that page
`n_tables_on_page`	`int32`	number of attached export tables; may be fewer than the tables visible on the page
`b_number`, `page_index`, `printed_page`		report id, 1-based scan index, original numeric printed-page value (nullable when regrouped values differ); exact export labels are in `signals.export_page_labels`
`source_collection`, `label_source`, `split`, `signals`		provenance (`label_source="table-export"`)

The original selection spans all 12 decades. Train/validation/test assignments remain grouped by report, with no report shared across splits. After the September 2026 alignment repair there are 2,520 / 302 / 327 rows respectively. Page counts increased because tables previously conflated by repeated printed-page numbers now occupy separate scans.

Alignment uses each table's content against every page's ALTO in its report; numeric canvas labels alone do not decide the match. The automated rule requires IDF-weighted token containment ≥0.70, a margin ≥0.10 over the next-best scan, and at least eight distinct content tokens. Two low-margin matches were retained after explicitly recorded agent visual checks. These scores measure agreement with source OCR, not the probability of a correct annotation. Limited image checks supplement the automated pass; the entire dataset has not been manually verified.

The alignment audit accounts for all 4,086 original table items: 3,529 retained on their original scan, 451 moved, and 106 unresolved. Unresolved items are excluded from the image/table config and preserved with their original CSV content in unresolved.jsonl. Table IDs, CSV strings and original report split assignments are conserved across the config and that side artifact. The default corpus is unchanged.

Target quality — silver, not gold. These tables were extracted by OCR/table recognition, not hand-transcribed. Known errors include incorrect digits, lost fractions, noisy captions and labels, flattened headers, shifted or missing cells, and inconsistent nil markers. Correcting the scan association does not correct those errors. Neither exact numeric scoring nor fuzzy text matching by itself accounts for this noise. Use an independently reviewed reference set for headline evaluation results. The original CSV strings, including any literal escaped line endings, are preserved; the repair script interprets those escapes only when parsing content for alignment.

Provenance & reproducibility

Images: Wellcome IIIF Presentation v2 API, keyed on each report's b-number. default uses native resolution; tables retains the existing approximate 800/1,000-pixel width convention, including for replacement scans. Actual dimensions (including IIIF rounding differences) are recorded in signals.image_size.
Text (default): the bulk Fulltext.zip corpus from wellcomelibrary.org/moh, joined on b-number.
Tables (tables): Wellcome's All_Report_Tables export (~275k machine-extracted tables). The original numeric-label resolver conflated duplicate and irregular printed page labels. The repair script performs report-wide ALTO content matching and preserves the original export payloads.
signals records the manifest and image service for each page. In tables it also records the canvas label, export-page labels, image dimensions, source dataset revision and alignment audit location. Per-table decisions and source checksums are in the audit.

Licence

Page images are CC-BY-NC 4.0 (per Wellcome's IIIF manifests; per-row license + signals.license_raw). The OCR text corpus is CC-BY 4.0; the extracted-tables export is CC-BY 4.0. Reports whose image licence was not open were excluded. Reuse is non-commercial, with attribution to Wellcome Collection.

licence	pages (`default`)
cc-by-nc	391,964

Source & attribution

Wellcome Collection, London's Pulse: Medical Officer of Health reports 1848–1972. Images: iiif.wellcomecollection.org · Text/tables: wellcomelibrary.org/moh.

Intended uses

Historical OCR/VLM evaluation, document layout analysis, page-type classification, and structured extraction from century-old public-health tables (disease incidence, mortality). The tables config provides candidate material for image→table extraction evaluation; reliable headline results require independently reviewed references.