README.md

metadata

pretty_name: Tweets Sample 2026 — Newspaper-Vocabulary Reference Corpus
language:
  - en
  - de
multilinguality:
  - multilingual
language_creators:
  - found
annotations_creators:
  - no-annotation
source_datasets:
  - original
size_categories:
  - 10M<n<100M
task_categories:
  - text-classification
  - text-retrieval
tags:
  - twitter
  - x
  - social-media
  - nitter
  - german
  - narrative-detection
  - reference-corpus
  - computational-social-science
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet

Tweets Sample 2026 — Newspaper-Vocabulary Reference Corpus

A large, deliberately untargeted sample of public posts from X/Twitter, collected via Nitter by sweeping a 65,689-term newspaper vocabulary rather than a topical keyword set.

It is built as a background / reference corpus: a baseline of "what was being said in general" against which a topically targeted collection can be contrasted. It is the reference arm of a narrative-detection study, not a curated dataset about any particular subject.

Two columns do not contain what their names say. quotes holds like counts and likes holds view counts. See Engagement columns are mislabeled before using any engagement metric.

At a glance

Rows (tweets)	75,845,834
Unique tweet IDs	75,845,834 (fully deduplicated — 1 row per tweet)
Unique authors	9,215,132 (≈ 8.2 tweets per author)
Date range	2006-03-24 → 2026-07-27 (99.81 % falls in 2026-01-01 → 2026-07-27)
Search terms swept	65,689
Distinct search_term values present	59,383
Size	25.7 GB, 73 Parquet shards (snappy)
Language mix (heuristic)	~52 % English, ~22 % German, remainder short/other
Splits	train only

Loading

The corpus is 25.7 GB; stream it unless you have the RAM to spare.

from datasets import load_dataset

# streaming — no full download, constant memory
ds = load_dataset("SinclairSchneider/tweets_sample_2026", split="train", streaming=True)
print(next(iter(ds)))

# or materialize the whole thing (needs ~25 GB disk + a lot of RAM)
ds = load_dataset("SinclairSchneider/tweets_sample_2026", split="train")

For analytical work, querying the Parquet files directly is far more efficient than going through datasets, because you can push column selection down to the file:

import duckdb

duckdb.sql("""
  SELECT search_term, count(*) AS n
  FROM 'hf://datasets/SinclairSchneider/tweets_sample_2026/data/*.parquet'
  GROUP BY 1 ORDER BY n DESC LIMIT 20
""").show()

The search vocabulary

search_terms.txt (shipped with this dataset) contains 65,689 lowercase, alphabetically sorted, single-token terms spanning aachen → übungsplatz. It is a general newspaper vocabulary — overwhelmingly ordinary words, not a political keyword list:

aachen, aal, aalglatt, aasgeier, abandon, abandoned, abarbeiten, abartig, abattoirs, abba, …
überziehend, überzieht, überzogen, üblich, übrig, übriggebliebener, übung, übungsplatz

Despite the file name (political_reference_newspapers.txt), the list is not topically political; the "political" label refers to the study it was assembled for. It mixes German and English vocabulary, which is why the corpus skews English (see Language composition).

Hits are very unevenly distributed across terms:

Percentile of terms	5th	25th	Median	75th	95th
Tweets returned	13	101	382	1,357	5,570

The highest-yield terms are English and mostly generic: leftie (63,980), statistic (58,708), bootlicker (48,482), screenings (47,040), stadion (46,015), oversold (44,747), preferable (42,918).

Data fields

Field	Type	Notes
search_term	string	The query that surfaced this tweet. Lossy — read the caveats below.
id	string	Tweet ID (numeric, stored as string). Unique across the corpus.
link	string	Permalink, e.g. https://twitter.com/{user}/status/{id}#m
text	string	Tweet body as rendered by Nitter
name	string	Author display name
username	string	Author handle, @-prefixed
profile_id	string	Author profile identifier
avatar	string	Author avatar URL
date	string	YYYY-MM-DD HH:MM:SS+00:00 (UTC). String, not a timestamp — lexically sortable.
is-retweet	bool	3,605,695 rows (4.75 %)
is-pinned	bool	Always false — never populated
external-link	string	Non-empty for only 215,350 rows (0.28 %); empty string otherwise
replying-to	list<string>	Handles replied to. Non-empty for 43,997,721 rows (58.01 %)
quoted-post	struct	Nested quoted tweet (date, link, text, user{…}, pictures, videos, gifs). The struct is always present, but its fields are null unless a quote exists — 6,992,385 rows (9.22 %) have one. Test quoted-post.link IS NOT NULL, not quoted-post IS NOT NULL.
comments	int64	Reply count
retweets	int64	Retweet count
quotes	int64	⚠️ Contains like counts
likes	int64	⚠️ Contains view counts
pictures	list<string>	Image URLs. Non-empty for 13,002,600 rows (17.14 %)
videos	list<null>	Always empty — video extraction never produced output; typed list<null>
gifs	list<string>	Non-empty for 913,666 rows (1.20 %)

Known issues and caveats

Engagement columns are mislabeled

The scraper reads the four tweet-stat elements of a Nitter tweet card by position and assigns them the names comments, retweets, quotes, likes. On the instances actually used, positions 3 and 4 are likes and views — there is no quote count in the markup — so the last two columns are shifted by one:

Column name	Actually contains	Median	99th pct	Max	% zero
comments	replies ✅	0	46	751,521	61.10 %
retweets	retweets ✅	0	205	1,339,671	75.09 %
quotes	⚠️ likes	2	1,176	4,228,611	31.14 %
likes	⚠️ views	89	51,433	1,539,096,983	0.71 %

The ratios make the shift unambiguous: the quotes column runs at 4.6× the retweet count, which is the canonical likes-to-retweets ratio (real quote counts are a fraction of retweets). The likes column runs at 99.7× retweets and 28.4× the quotes column, with a maximum of 1.54 billion — impossible for likes, entirely normal for views. Only 0.71 % of rows are zero, consistent with X displaying a view count on nearly every post.

Rename on load:

df = df.rename(columns={"comments": "replies", "quotes": "likes", "likes": "views"})

A true quote count does not exist in this dataset.

search_term is lossy in two ways

1. It is one arbitrary matching term, not all of them. A tweet containing several vocabulary words was returned by several queries, but deduplication kept only the first row encountered in filesystem-glob order. search_term is therefore a term that matched, chosen arbitrarily — never treat it as the complete set of matching terms, and do not use it as a label or as a frequency estimate for that term in the corpus.
2. Long terms are truncated to their last 20 characters. The value is recovered from a filename built with term[-20:], so 455 of the 65,689 terms (0.69 %) appear clipped — entscheidungskompetenz → tscheidungskompetenz, nachfolgeorganisation → achfolgeorganisation. Truncation can in principle also merge two distinct long terms that share a suffix.

Dates fall outside the requested window

Although the sweep requested since=2026-01-01, 144,215 rows (0.19 %) predate 2026, reaching back to 2006-03-24 — the search backend does not honour the date bound strictly. Filter on date yourself if you need a clean window.

Period	Tweets
Before 2025	77,052
2025	67,163
2026-01	6,521,317
2026-02	6,565,188
2026-03	8,166,263
2026-04	9,419,069
2026-05	16,926,040
2026-06	16,553,381
2026-07 (to the 27th)	11,550,361

The 2026 month-over-month growth is a collection artefact, not a signal about platform activity — scraping ran forward in time and later months had more instance capacity available. Do not read the monthly curve as a volume trend.

Language composition

No language filter was applied, and the vocabulary itself is mixed, so the corpus is majority English despite being assembled from a German newspaper vocabulary. A stopword heuristic over a 6,639,818-tweet sample:

	Share
English markers only	52.3 %
German markers only	22.2 %
Both	3.8 %
Neither (short texts, other languages)	21.7 %

This is a coarse heuristic, not a trained language ID. Run a proper detector (fastText lid.176, GlotLID) if language is load-bearing for your work.

Sampling bias

This is not a random sample of X, and no weighting will make it one:

• Coverage is bounded by the vocabulary. A tweet containing none of the 65,689 terms cannot appear — which biases against very short posts, emoji-only posts, other languages, and heavy slang.
• Nitter search returns what the search backend chooses to return, with its own opaque ranking and recall limits. Unlimited --max_tweets does not mean exhaustive retrieval.
• Instance availability varied over the collection period, so coverage is uneven across time and across terms — 6,306 of the 65,689 vocabulary terms (9.6 %) yielded nothing at all, and it is not determinable after the fact whether a given term genuinely had no matches or was simply never served successfully.
• Engagement-related selection is unknown: high-visibility tweets are plausibly over-represented.

Other

• Content reflects the moment of scraping. Tweets later deleted, edited, or made private are still here; counts are frozen at scrape time and no longer match the platform.
• Retweets (4.75 %) duplicate the text of the original post — filter is-retweet for text work.
• date is a string. Cast it before doing time arithmetic.

Intended use

Suited to: building background/reference distributions for narrative and framing detection, term-frequency baselines, corpus-linguistic study of platform language, retrieval and topic-model development, and pretraining or domain adaptation of social-media models.

Not suited to: measuring the prevalence of anything on X (the sampling frame forbids it), tracking trends over time (the monthly curve is an artefact), studying individuals or accounts, or any engagement analysis that has not first corrected the mislabeled columns.