Datasets:
The dataset viewer is not available for this split.
Error code: FeaturesError
Exception: ArrowInvalid
Message: Schema at index 1 was different:
band: int64
prompts_passing: int64
of: int64
prompts: string
vs
prompt_id: string
n_machine_in_word_corridor: int64
chars_lo: int64
chars_hi: int64
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/split/first_rows.py", line 244, in compute_first_rows_from_streaming_response
iterable_dataset = iterable_dataset._resolve_features()
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 4523, in _resolve_features
features = _infer_features_from_batch(self.with_format(None)._head())
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2768, in _head
return next(iter(self.iter(batch_size=n)))
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2972, in iter
for key, pa_table in ex_iterable.iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2483, in _iter_arrow
yield from self.ex_iterable._iter_arrow()
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 591, in _iter_arrow
yield new_key, pa.Table.from_batches(chunks_buffer)
~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^
File "pyarrow/table.pxi", line 5012, in pyarrow.lib.Table.from_batches
File "pyarrow/error.pxi", line 155, in pyarrow.lib.pyarrow_internal_check_status
File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
raise convert_status(status)
pyarrow.lib.ArrowInvalid: Schema at index 1 was different:
band: int64
prompts_passing: int64
of: int64
prompts: string
vs
prompt_id: string
n_machine_in_word_corridor: int64
chars_lo: int64
chars_hi: int64Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Equal Scores, Unequal Choices — data
Data for the paper Equal Scores, Unequal Choices: LLM Judges Favour Machine-Written Essays. LLM judges score (1–6) and choose between student essays from PERSUADE 2.0 and ASAP 2.0, which trained raters scored on the same holistic scale, and machine-written essays for the same prompts.
The accompanying code (https://anonymous.4open.science/r/essay-judges-2C42/) reads this directory as its data/ folder: make paper rebuilds every table,
figure and number of the paper from results/, and make analyses recomputes results/ from the
judge answers in judgments/.
Contents
| Path | Rows | What |
|---|---|---|
prompts/ |
15 | The writing prompts: 8 PERSUADE 2.0 independent prompts and 7 ASAP 2.0 source-based prompts with their source texts. |
human/ |
37,848 | Student essays with the trained raters' holistic score (quality_score, 1–6). |
essays/raw/ |
840 | Machine-written essays (14 models, default generation), as generated. |
essays/normalized/ |
38,688 | Human and machine essays after the same form-only normalization (text_normalized). |
essays/variants/*.parquet |
Rewritten and degraded essays, keyed to their source essay (source_id): g2, g2b (machine essays rewritten in a student's style), g3 (human essays polished by a model), d1 (machine essays with student-rate surface errors), d2 (content weakened, with surface errors), d3 (d2 with paragraphs shuffled), lc (human essays with spelling, grammar and punctuation corrected), nt (essays rewritten in plain standard English). |
|
judgments/<name>/calls.jsonl |
150,072 | Every judge answer, one JSON object per request (see below). |
results/ |
Every analysis output the paper is built from. |
Parquet tables in prompts/, human/ and essays/ carry a manifest.json (row counts, schema,
checksums).
Judge answers
Each line of judgments/<name>/calls.jsonl is one request to one judge:
_key:experiment(arm code; seesrc/essay_judges/judging/units.pyin the code),judge(model),kind(pointfor a 1–6 score,choicefor picking the best of several essays),pool(the essay or pool id) andorder(presentation order);candidates: essay ids in the order shown;display_ids: the opaque ids the judge saw;labels: authorship labels shown, if any;condition: the prompt condition (e.g. a remedy);prompt_id,prompt(the full request text),prompt_sha256;raw(the judge's reply),value(the parsed score or chosen id),status(validor the reason the answer is invalid),detail,error,finish_reason;response_id: an anonymised identifier of the response, distinct for every model call.
| Name | Content |
|---|---|
pointwise_main, pointwise_pool_judges, pointwise_flagships, pointwise_gemma_26b, pointwise_size_ladders, pointwise_qwen, pointwise_haiku, pointwise_reasoning_low |
1–6 scores of human and machine essays by each judge panel |
pointwise_rewrites, pointwise_mechanism_items |
scores of rewritten, corrected and neutrally rewritten essays |
pointwise_repeats |
further independent repeats of identical scoring requests (test-retest noise) |
pools_six_and_labels, pools_six_flagships |
one-best choice among six essays, with and without authorship labels |
equal_score_degraded, equal_score_pairs, equal_score_controls, equal_score_gap_controls, equal_score_size_ladders |
the equal-score test: human–machine pairs the judge scored equally, same-class control pairs, degraded pairs, harder and rubric pools |
pools_k, odd_one_out |
pools of up to ten candidates and all their pairs |
remedies, reasoning_pairs |
prompt-level remedies; judges at low reasoning effort |
neutral_rewrite_and_error_dose, neutral_rewrite_and_error_dose_sonnet |
neutrally rewritten pools; essays as written, lightly corrected and polished |
size_ladder_pairs_qwen, size_ladder_pairs_haiku |
the equal-score pairs for further models of the size ladders |
Not included
- Per-essay demographics of PERSUADE 2.0 (English-learner and economic status), used for the group gaps at equal rater band. Only aggregates are reported; the code fetches the per-essay fields from the original release.
- BAWE essays (British Academic Written English corpus), whose terms do not allow redistribution.
They supply part of the reference vocabulary of one feature (misspellings) of the
machine-likeness classifier; the computed features are in
results/.
Sources and licence
Human essays and prompts come from:
- PERSUADE 2.0 — Crossley et al. (2024), A large-scale corpus for assessing written argumentation: PERSUADE 2.0, Assessing Writing 61. Licence: CC BY-NC-SA 4.0.
- ASAP 2.0 — Crossley et al. (2025), A large-scale corpus for assessing source-based writing quality: ASAP 2.0, Assessing Writing 65. Part of its essays also appear in PERSUADE 2.0.
This dataset is released under CC BY-NC-SA 4.0, the licence of PERSUADE 2.0, which the human essays, the prompts and the judge requests (which contain essay texts) inherit. Machine-written essays and judge answers were produced through the providers' APIs for research use.
- Downloads last month
- 41