Cambridge 2026 · UK headlines · data check

AG vs UK headline length

How long are Stage-1 AG titles compared with BBC UK headline proxies? Length mismatch is one reason a model trained on AG may struggle on UK text before domain adaptation.

Mean words: AG 13.4 vs UK 16.0
Mean chars: AG 83.9 vs UK 94.6
Exactly 16 words: AG 59% · UK 99%

Summary stats

AG = train + val titles (n = 30,000). UK = all prepared BBC rows after dropping entertainment (n = 1,839). Whitespace-token word counts.

Corpus n Mean words Median Std P25–P75 Mean chars Char median
AG News 30,000 13.41 16 4.33 10–16 83.93 90
BBC UK 1,839 15.97 16 0.39 16–16 94.63 94
Why UK is so peaked at 16 words: BBC rows are often full articles, so prepare_data.py builds a headline proxy (first sentence, else first ~16 tokens). That caps almost every UK example at 16 words — character length still varies.

Word-length distribution

Percent of headlines at each word count (normalized so different corpus sizes are comparable).

Character-length distribution

Even with a shared word cap, UK proxies are a bit longer in characters on average (denser around 80–110 chars).

Mean length by label

Mean word count per unified topic label.

Label AG mean words UK mean words AG mean chars UK mean chars
business 13.38 16.00 84.2 94.0
sport 13.40 15.95 81.0 96.7
tech 13.75 15.94 87.6 92.9
politics 13.12 16.00 82.9 94.6