sjin4861/sail-asap2-lopo

SAIL: ASAP 2.0 leave-one-prompt-out splits and simulator-written reference pools

Everything a reader needs to reproduce the paper's inputs exactly, and the one artefact that is new: the essays the simulator wrote for prompts it had never read.

Contents

path	what it is
corpus/processed.tsv	ASAP 2.0, both release CSVs pooled and keyed by prompt id. 24,728 rows. Essays are verbatim full_text; nothing is normalised and nothing is excluded.
corpus/source_texts/	The passage set each prompt supplies, as plain text.
references/fold_k/pool.tsv	The synthetic pool. 300 essays written by the fold's simulator for the held-out prompt, 50 at each score 1--6.
references/base/fold_k/pool.tsv	The same protocol run on the pretrained model with no simulator fine-tuning.
splits/	The K=5 folds as trained: simulator/fold_k/ and scorer/fold_k/ with references already attached.
splits_shot{7,9,11}/	The same folds at larger reference counts, for the K ablation.
oracle_splits/	The all-prompt (non-LOPO) split behind the reporting-only judge.
predictions/	Per-essay scorer outputs behind the reported numbers.

The splits

Fold k holds out prompt k+1 entirely; the other six prompts are shuffled per prompt with seed 42+k and split 90/10 into train and validation, so the validation half carries all six prompts in their corpus proportions. The simulator and the scorer use the same partition.

fold	target	train	valid	test
0	P1	20,413	2,269	2,046
1	P2	20,491	2,278	1,959
2	P3	17,859	1,986	4,883
3	P4	16,701	1,857	6,170
4	P5	18,222	2,026	4,480
5	P6	20,297	2,256	2,175
6	P7	19,541	2,172	3,015

The release's own train/test column is deliberately ignored --- it leaves two prompts with no test rows at all --- but preserved per row as release_split.

Reading the pools

Each row of a pool.tsv carries requested_score and label_source='requested'. These are the scores the essays were asked to earn, not human ratings and not read-back estimates. An evaluation that treats them as ground truth is measuring something else.

Coverage is uniform by construction --- 50 essays in every band --- because the held-out prompt's real score distribution is target label data, and shaping the pool with it would be the leak leave-one-prompt-out exists to prevent. The empirical distribution enters one stage later, when references are drawn for a target essay: the score band is drawn from the six source prompts' distribution and the essay uniformly from that band.

Every generated essay passed a validity filter whose thresholds are read off the corpus: at least 150 words (the shortest real essay of every prompt is 150 or 151), a distinct- 4-gram ratio of at least 0.70 (only 7 of 4,000 sampled real essays fall below), no echo of the simulator's own instruction headers, and a sequence that stopped on an end token rather than the token budget. Rejected samples were resampled, never repaired.

The reference draws are nested

References are drawn one at a time from a single per-row seeded generator, so the five references of a K=5 row are five of the eleven a K=11 row carries. Verified on the written files: all four splits of all seven folds, 197,824 rows at each K, no exception, every reference distinct, never the target itself, always from the target's own prompt. The K ablation therefore varies only how many references, not which.

Licence and citation

ASAP 2.0 is released CC BY 4.0 at scrosseye/ASAP_2.0; the derived splits here carry the same licence. Please cite Crossley et al., A large-scale corpus for assessing source-based writing quality: ASAP 2.0, Assessing Writing 65 (2025).