You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset contains scanned Texas county land records (public records, but they contain personal names and addresses). Access is granted manually. Describe your intended use below.
Log in or Sign Up to review the conditions and access this dataset content.
Madison County, Texas - Land Records
Recorded land-records documents (deeds, mortgages, liens, releases, plats, etc.) from the Madison County Clerk's free online
index: i2i.uslandrecords.com/TX/Madison. Each document is the original scanned page images, unmodified (JPEG, 2550 px
wide, exactly as served), the docTR word-level OCR output for each page, the clerk index row (parties, legal description,
references) and per-page OCR text.
These are public records from the Madison County Clerk. The dataset is gated (manual approval) because the records contain personal names and property details; request access on the dataset page.
Contents (updated 2026-10-19T06:17:56)
| Documents | 13,354 |
| Pages (images) | 51,428 |
| Shards (data + raw) | 50 |
| Size (data + raw) | 44.0 GB |
| Documents with OCR sidecar (ocr/) | 218 |
| Filing years covered | 1840-2026 |
Only whole records are included (every page image downloaded and checked against the clerk's page count, parties complete).
A document is uploaded as soon as it is completely downloaded, before OCR (raw/); its OCR is added afterwards as a separate
sidecar record (ocr/). New documents are added incrementally; shards are append-only and never rewritten, so a document's OCR
is always in a later shard than its images. Documents uploaded earlier (before 10-17) are in data/ with images and OCR together.
Layout
raw/raw-NNNNNN.tar- WebDataset shards (~1 GB), current format. One record per document, grouped by key:<key>.json- clerk index + detail metadata (see below; no OCR)<key>.pNNN.jpg- original page image (page NNN, byte-identical to the source)
ocr/ocr-NNNNNN.tar- OCR sidecars for documents already inraw/(separate shards, uploaded after the document has been read):<key>.ocr.json-{key, doc_no, ocr_at, pages: [{page, ocr, text}]}<key>.pNNN.ocr.json.gz- docTR word boxes + text for page NNN Join raw and OCR records onkey.
data/shard-NNNNNN.tar- old combined shards (image + OCR + metadata per document, same member names as above, OCR inside<key>.jsonpage_data).manifest/raw.jsonl,manifest/ocr.jsonl,manifest/manifest.jsonl- one line per uploaded document in the matching track:key,doc_no,shard,pages,bytes,year.
Key
key = madison_<doc_no> where every character outside [A-Za-z0-9-] in doc_no is replaced by _<hex>_
(for example 968@1955-03-14 -> madison_968_40_1955-03-14). The original doc_no is in the JSON; the mapping is reversible.
JSON fields
key, doc_no, county, county_slug, file_date (M/D/YYYY), rec_time, instr_date, doc_type, pages, img_pages,
book_page, legal, grantors (list), grantees (list), refs, n_parties, parties_incomplete, text_status,
detail_raw (clerk detail page text), harvested_at, ocr_at, downloaded_at (UTC, from image files), source,
page_data (list of {page, image, ocr, image_bytes, image_md5, ocr_bytes, text}).
Note: very old records sometimes have an empty/partial party list (parties_incomplete); OCR text is machine output and
contains errors.
Loading
import webdataset as wds
ds = wds.WebDataset("data/shard-{000000..000003}.tar")
License / terms
Source documents are public records. No license is asserted over the underlying records; the compilation, OCR output and metadata are provided for research and title-search use under the access conditions above.
- Downloads last month
- 250