Download: votes.csv · items.json · scores.json · browse all 60 pairs

Do LLMs have taste? — data

Everything behind the post Do LLMs have taste?.

The experiment

60 pairs. Each pair puts one passage written by Claude against one 150-word excerpt from a public-domain novelist or poet, matched by form: prose against prose, verse against verse. 45 distinct passages — 15 machine, 30 human.

Each pair was judged by 5 models in both presentation orders. That is 2 judgments per judge per pair, 120 per judge, 600 in all.

Judges were told the passages were literary fiction and poetry, and asked which was the better piece of literary writing. They saw no labels, no authors and no context, and were never told a machine was involved. A fresh judge session handled each batch, and the two orders of a pair always went to different sessions.

A pair is decided only if both orders agree. If they disagree, that is position bias and it is recorded as a tie, not as one vote each way. This is why the win rates below have different denominators.

judge chose the machine ties
Claude Opus 42 of 48 decided (88%) 12
DeepSeek V4 Flash 34 of 36 (94%) 24
GPT-5.6 Luna 54 of 56 (96%) 4
GPT-5.6 Luna Pro 56 of 57 (98%) 3
Claude Haiku 4.5 58 of 59 (98%) 1

A second framing was also run — the same 60 pairs under a plain "which is the better piece of writing?", with no mention of literature. It changed nothing (Opus 88% under both, no decided pair flipped), so only the literary arm is published here.

votes.csv — one row per judgment (600 rows)

column meaning
judge the model that made this judgment
pair_id which pair; 60 distinct values
stratum lit:machine_lit (prose) or verse:machine_verse (verse)
order ab or ba — which side of the pair was shown first
x_id, y_id the passages shown as X and Y, in that position
choice X or Y, the letter the judge answered
winner_id the passage that letter refers to
winner_kind machine or human — the column most people want
reason the one sentence the judge gave, verbatim

X and Y are positions, not identities: the same passage is X in one row and Y in another. Always read winner_id / winner_kind, never the letter.

items.json — the 45 passages (full text)

id, kind (machine/human), form (prose/verse), author, title, source_url, words, text. Human passages are public-domain, via Project Gutenberg. Machine passages were written to a one-line topic prompt with no style instruction.

⚠ One byline is wrong and is left as the judges saw it: the passage filed under Coleridge's Biographia Literaria is Wordsworth's "Yew-Trees", quoted inside it. Several judges named Wordsworth correctly. It carries an author_note.

scores.json — resolved verdicts

judges holds each judge's totals. pairs holds one entry per pair (60), with the machine and human passage ids and each judge's resolved verdict — winner_id: null means the two orders disagreed.

What this data cannot tell you

The human side is excerpts and the machine side is whole pieces, and judges sometimes penalise the excerpt for being one. "Which is the better piece of literary writing?" is an evaluative frame that invites a rubric, and a different question might get different answers. 60 pairs is small, and after ties are set aside some judges rest on fewer than 40 decided pairs.