A dated, deduplicated corpus of technology news, trade press and vendor publications: 6,489,215 articles from 1,383 hosts, full extracted text (6,674,439 rows with the outliers).

Config	Rows	What it is
`all` (default)	6,489,215	`dated` and `undated`
`window`	4,135,904	published 2023-07 to 2026-09, the three-year window the corpus was built for
`dated`	4,502,047	every article with a publication date, 2020-01 to 2026-09
`undated`	1,987,168	pages with no evidenced publication date
`outliers`	185,224	pages whose dates are not trustworthy as publication dates, kept apart (see below)

Load it

from datasets import load_dataset

ds = load_dataset("IDONTCARELETMEIN/tech-trade-press", "window", split="train", streaming=True)
for row in ds.filter(lambda r: r["host"] == "tomshardware.com").take(5):
    print(row["published"], row["title"])

DuckDB reads only the columns and row groups it needs:

INSTALL httpfs; LOAD httpfs;
CREATE SECRET hf (TYPE huggingface, PROVIDER credential_chain);  -- picks up your hf auth login token
SELECT date_trunc('month', published) AS month,
       count(*) FILTER (WHERE text ILIKE '%solid-state batter%') AS n,
       count(*) AS total
FROM 'hf://datasets/IDONTCARELETMEIN/tech-trade-press/articles/dated/*/*.parquet'
WHERE lang = 'en' AND published >= DATE '2023-07-01'
  AND host_first <= '2023-10'   -- only hosts covered across the whole window
GROUP BY 1 ORDER BY 1;

One month to disk:

hf download IDONTCARELETMEIN/tech-trade-press --repo-type dataset \
  --include "articles/dated/2025-03/*" --local-dir ./f50d

Polars, lazily:

import polars as pl
lf = pl.scan_parquet("hf://datasets/IDONTCARELETMEIN/tech-trade-press/articles/dated/*/*.parquet")
lf.filter(pl.col("topics").list.contains("ai")).group_by("host").len().sort("len", descending=True).head(20).collect()

What is in it

Counts below are for all (dated and undated); outliers is not included.

Rows by publication year:

Year	Rows
2020	77,640
2021	90,920
2022	131,956
2023	334,647
2024	959,188
2025	1,088,480
2026	1,819,216
undated	1,987,168

Languages:

Language	Rows	Share
en	4,746,830	73.1%
ru	744,198	11.5%
zh	318,722	4.9%
ko	248,366	3.8%
ja	164,717	2.5%
fr	136,057	2.1%
other	130,325	2.0%

Where the date came from:

date_source	Rows	Share
meta	2,637,532	40.6%
(undated)	1,987,168	30.6%
jsonld	1,546,981	23.8%
url	204,206	3.1%
time	103,728	1.6%

Source format:

format	Rows	Share
html	6,479,423	99.8%

Topics (a row can carry several):

Topic	Rows	Share
technology	2,176,556	33.5%
investment	1,499,852	23.1%
domain	1,345,391	20.7%
money	1,068,662	16.5%
ai	989,733	15.3%
innovation	613,439	9.5%

relevant is true on 2,974,469 rows (45.8%). Text length: median 3,113 characters, 90th percentile 11,370.

Layout

articles/dated/<YYYY-MM>/NNNN.parquet   by publication month, 2020-01 to 2026-09
articles/undated/NNNN.parquet           no evidenced publication date
articles/outliers/NNNN.parquet          dates kept, but not trustworthy; see date_flag

Files hold up to 100,000 rows in row groups of 5,000, zstd-compressed.

Columns

Column	Meaning
`id`	Row id
`url`, `host`	Source page and its host
`lang`	Language of the text, py3langid on the first 2,000 characters; rare labels are often misdetections
`published`	Publication date, or null in `undated`; in `outliers`, the date the page gave
`date_source`	Where the date came from (see the table above); empty when undated
`format`	Source format: `html`, `pdf`, `aem`, `xml`
`fetched_at`	When the page was fetched (UTC)
`title`, `text`	Extracted title and main text
`chars`, `words`	Length of `text`
`text_sha1`	SHA-1 of `text`; the corpus is deduplicated on it and on `url`
`relevant`	Heuristic flag: the page touches technology, investment or innovation vocabulary
`topics`	Heuristic topic labels, e.g. `ai`, `technology`, `investment`, `money`
`tags`	Reserved; empty in this release
`money`	Candidate sentences with amounts, funding rounds, revenue or deal terms: `kinds`, `amount`, `sentence`
`deliverables`	Candidate sentences naming launches, releases, deployments, contracts: `trigger`, `sentence`
`date_flag`	Null, or in `outliers` why the date is not trusted: `bunched_host`, `stamped_day`
`host_first`	First month (`YYYY-MM`) with a trusted dated row from this row's host; `2020-01` means at or before it. Null when the host has none

One row, with text cut short:

{
  "id": 2811374,
  "url": "https://europeanspaceflight.com/uk-based-hyperspectral-data-analytics-startup-messium-raises-3-3m/",
  "host": "europeanspaceflight.com",
  "lang": "en",
  "published": "2025-09-15",
  "date_source": "meta",
  "format": "html",
  "fetched_at": "2026-09-26 21:25:00+00:00",
  "title": "UK-Based Hyperspectral Data Analytics Startup Messium Raises £3.3M",
  "text": "UK-based startup Messium has raised £3.3 million in seed funding to accelerate the rollout of its agricultural analytics platform, which uses AI to analyse hyperspectral satellite data to track nitrogen levels in crops. The analysis can then be used by farmers to more efficiently fertilise crops, mi...",
  "chars": 2605,
  "words": 400,
  "text_sha1": "53532d7b7781af40a27b3b7b188f12acc9507cfc",
  "relevant": true,
  "topics": [
    "ai",
    "domain",
    "innovation",
    "investment",
    "money",
    "technology"
  ],
  "tags": [],
  "money": [
    {
      "kinds": [
        "amount",
        "round",
        "funding"
      ],
      "amount": "£3.3 million",
      "sentence": "UK-based startup Messium has raised £3.3 million in seed funding to accelerate the rollout of its agricultural analytics platform, which uses AI to analyse hyperspectral satellite data to track nitrogen levels in crops."
    }
  ],
  "deliverables": [
    {
      "trigger": "rollout",
      "sentence": "UK-based startup Messium has raised £3.3 million in seed funding to accelerate the rollout of its agricultural analytics platform, which uses AI to analyse hyperspectral satellite data to track nitrogen levels in crops."
    }
  ],
  "date_flag": null,
  "host_first": "2022-02"
}

How it was built

Hosts. The seed was 206 curated hosts across six domains (fintech, energy, medtech, logistics, industry, IT). Discovery passes added sources in sixteen further fields (space, medicine, infrastructure, chemistry, avionics and others) and Chinese-language trade press, each host checked by fetching it.
URLs. Every host's sitemaps were read (its listing pages, for hosts without one), and everything inside the three-year window was crawled until the host refused. Pages outside the window were kept when the crawl reached them.
Text. The main text of each page, without navigation, ads or boilerplate. PDF, XML and AEM JSON bodies are converted to HTML first and extracted the same way.
Dates. From the page itself: meta tags, JSON-LD, the URL, <time> elements, or the document's own metadata. date_source names which. A page with no date evidence is undated.
Pruning. Every page was stored from a successful response. A response other than a 200, a challenge page, a redirect to a homepage, an empty or near-empty extraction, a JSON or non-HTML page with no convertible body, and gibberish were all dropped.
Deduplication. Exact duplicates by text_sha1, keeping the first; each url appears once.
Tags. money and deliverables are regex candidates; relevant and topics are keyword rules. None of them was reviewed by hand.
Cutoff. Pages with a publication date before 2020-01-01 were dropped. Undated pages are kept, in undated.
Outliers. Rows whose dates cannot be trusted as publication dates are moved to outliers with their date kept and a date_flag:
- bunched_host: every dated row of a host with at least half of its dated rows inside three consecutive days, or, for a host with under 20 dated rows, 90% on one day. These are archives stamped with one date, and hosts whose rows all fall in the last days before the crawl ended.
- stamped_day: a host's day with at least 200 rows, at least 20 times that host's median day and at least 5 times the average of the 7 days either side of it, unless the page URL carries that date. A busy stretch of real news days is not isolated and stays dated. These are archive pages stamped with a record-update date (project reports, publication lists, case pages).
In this release: bunched_host 81,620 rows from 401 hosts; stamped_day 103,604 rows from 30 hosts.

Uses

Built to detect emerging technologies and trends early, by tracking a term's monthly share of coverage and how it changes. It also serves retrieval and search over tech press, and training and evaluating classifiers on the heuristic tags.

Known limits

Coverage is a curated host list, not the open web, and monthly volume is uneven. From 2023-07 to 2026-09 the median month has 88,120 articles, the middle half 72,748 to 103,571. 2023-10 to 2025-12 runs at 56k to 104k a month. 2023-07 to 2023-09 are thin (11k to 42k): they sit at the start of the three-year crawl window. 2026 rises from 118k in 2026-01 to 242k in 2026-09, so its 9 months hold 44% of the window. Hosts per month: 456 in 2023-10, 529 in 2026-09. The ramp is more pages per host in recent months, not new sources: sites list recent pages in their sitemaps far more fully than old ones. Compare a term's share of each month's total, not raw counts.
Hosts join the corpus over time. 67 hosts have their first dated row in 2026 and hold 21% of 2026's rows. 600 hosts have host_first at or before 2023-10; they hold 71% of the window. A host that appears in a given month makes the terms it covers look as if they start rising then. For trends, keep to hosts present across the window with host_first, for example host_first <= '2023-10'.
Dates before 2023-07 come from the archives of hosts crawled for the window. They are a baseline, not a balanced sample of those years.
31% of the rows have no publication date. They are full articles, useful for search and retrieval, not for time series.
Some pages are not articles (listings, tag pages, product pages). words and relevant help filter them.
On some hosts the same block repeats on every page: a cookie or privacy notice, a disclaimer, a subscription pitch or a sidebar, at the start or end of text. On 235 hosts at least a fifth of the rows share the same opening or closing text, about 1.6 million rows in all. A sidebar holds what the site showed when it was fetched, so a term that appears only there shows up on every page of that host, whatever its date. Before counting, strip the text that most of a host's rows share at the start or end of text.
A few paywalled hosts yield only a lead.
relevant, topics, money and deliverables are heuristic, not labels.

Rights

Article text belongs to its publishers: title, text, and the sentence fields of money and deliverables.
The dataset's own fields (url, host, lang, published, date_source, format, fetched_at, chars, words, text_sha1, relevant, topics, money.kinds, money.amount, deliverables.trigger, date_flag, host_first) are released under CC BY 4.0.
Every page was stored from a successful response. No block, challenge or error page is in the set.

Downloads last month: 104

Total file size:

14.4 GB