Datasets:
id stringlengths 3 3 | prompt stringlengths 50 80 | source stringclasses 1
value | license stringclasses 1
value | workflow_batch int64 0 24 | workflow_position int64 0 3 | used_in_completed_runs bool 1
class |
|---|---|---|---|---|---|---|
025 | He folded the map along the old creases and put it back unopened.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 9 | 0 | true |
075 | A widow watered a plant that had already dropped half its leaves.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 6 | 0 | true |
031 | A student underlined a sentence and then stared at it as if it might change.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 16 | 3 | true |
076 | The ticket office light stayed on after the last train had gone.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 8 | 2 | true |
014 | The night bus sighed at the stop and took no one.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 12 | 0 | true |
039 | A dog barked at a bicycle that had already gone around the corner.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 10 | 2 | true |
020 | The tailor left a pin in the cuff and wrote a note about Thursday.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 19 | 2 | true |
038 | The bakery's second tray came out paler than the first and they sold it anyway.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 9 | 1 | true |
100 | The chapel vestry key was under a brick that everyone already knew about.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 22 | 0 | true |
006 | The newsagent stacked magazines that nobody had asked for since June.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 3 | 2 | true |
023 | A farmer leaned on the gate and named the cow that would not come.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 17 | 1 | true |
021 | She boiled the kettle twice because the first whistle felt too early.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 15 | 2 | true |
012 | The river path was closed; someone had tied a ribbon to the gate.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 16 | 1 | true |
010 | The choir left their coats on the pews and the nave smelled of wool.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 23 | 3 | true |
015 | A boy practised a salute in the reflection of a dark shop window.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 10 | 0 | true |
084 | The canteen soup had a skin on it and they served it anyway.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 4 | 1 | true |
058 | The violin case on the empty seat kept its shape after the player left.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 20 | 2 | true |
007 | A cat watched the fishmonger from the dry side of the doorway.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 16 | 2 | true |
095 | A mechanic named the rattle before he opened the bonnet.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 2 | 2 | true |
071 | A teacher collected homework that smelled faintly of toast.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 0 | 3 | true |
008 | The lift stuck between floors long enough for two strangers to stop pretending.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 21 | 0 | true |
085 | She hung the washing under the porch because the line had snapped.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 1 | 1 | true |
051 | A nurse wrote the time in a margin and underlined it once.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 19 | 3 | true |
018 | The harbour office clock was six minutes fast and nobody reset it.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 23 | 0 | true |
035 | A courier checked the name twice on a door that had three bells.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 13 | 1 | true |
029 | She left the porch light on for a guest who had sent a later message.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 24 | 0 | true |
028 | The canal lock gates met with a sound like a book closing.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 21 | 3 | true |
068 | The dairy left bottles on the step and took back the ones from Tuesday.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 18 | 3 | true |
083 | A girl practised her scales behind a door that did not quite close.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 13 | 3 | true |
013 | She counted the remaining stamps as if they were a kind of money.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 4 | 0 | true |
042 | The lighthouse keepers' cottage had a new satellite dish and an old broom.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 9 | 2 | true |
099 | A student missed the last bus and sat on the kerb with a warm bag of chips.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 22 | 3 | true |
062 | The night market packed lanterns into crates that still smelled of oil.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 7 | 1 | true |
052 | The pier's closed stall had a sign about ice cream and a padlock from winter.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 5 | 3 | true |
004 | Rain found the hole in the market awning and made a dark coin on the table.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 11 | 0 | true |
065 | He tightened the jar lid with a cloth because his hands were wet.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 15 | 1 | true |
059 | A clerk dated a form with yesterday's stamp and did not notice.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 14 | 3 | true |
056 | The boarding house stairs creaked on the same step for every guest.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 2 | 0 | true |
040 | The hotel corridor carpet showed a path from the ice machine to the stairs.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 11 | 3 | true |
022 | The cinema's last showing ended and the usher swept popcorn into a quiet pile.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 6 | 1 | true |
079 | A barber swept hair into a pile that looked like nothing belonging to anyone.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 7 | 3 | true |
048 | The railway café served tea in cups that remembered a company name.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 9 | 3 | true |
088 | The stone trough held rainwater and a single floating leaf.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 6 | 2 | true |
092 | The greenhouse tap dripped onto a tray of seedlings in even time.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 24 | 1 | true |
024 | The attic window leaked onto a box of photographs that had already curled.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 20 | 3 | true |
016 | The greenhouse glass ticked as it cooled after the sun left.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 7 | 2 | true |
026 | The schoolyard bell rang for a drill that everyone already knew.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 3 | 0 | true |
096 | The village noticeboard held a concert that had already happened.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 10 | 1 | true |
046 | The quarry pond held the sky in a still rectangle after the wind died.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 13 | 2 | true |
091 | A woman buttoned a coat that had belonged to someone taller.
| https://github.com/jensabrahamsson/text-watermark-laboratory/tree/989064945295141dbc85b8acbc3a9026f7827a83/experiments/2026-09-01-pair-100x4 | mit | 23 | 1 | true |
WONI-CrossBench
WONI — Watermark Observatory for Networked Integrity is a collection of benchmark inputs and evaluation artifacts for language-model watermarking research.
Author: Euisuh Jeong
Affiliation: Artificial Intelligence Based Technology Company, Republic of Korea Air Force
WONI-CrossBench provides versioned prose prompts and structured-output tasks
for watermark evaluation. Release 2026-09-23-public-v3 (public-v3).
Counts and file hashes: MANIFEST.json.
| Configuration | Released records | Interpretation |
|---|---|---|
cross-scheme |
100 prompts: 50 train, 25 calibration, 25 test | Recorded exploratory reader-study split; all 100 were reused by the detector-conformance workflow |
structured-output |
450 tasks: 58 development, 387 held-out, 5 excluded schema siblings | Frozen allocation; this input-only release provides no completed held-out output matrix |
Sources: reader split,
workflow plan,
exposure boundary.
The 25 workflow batches contain four prompts each. workflow_batch and
workflow_position are zero-based. Engineering prompts in the plan are not
additional study prompts. “Held-out” denotes the frozen schema-group allocation,
not a claim that inference has never been attempted. The exposure boundary is
historical, not a live execution-status report.
Public-content boundary
One task, o60871, requires local reconstruction for exact replay. An example
field in its upstream JSON schema contains a third-party telemetry endpoint
with userinfo. Its validity was not tested. All six occurrences across three
files are replaced by an explicit marker. No task, split, reference answer or
recorded score was removed. The directory name original/ is historical: two
of its task files now contain that documented substitution and must not be
claimed byte-identical until reconstructed. PUBLICATION.json
records affected paths, original hashes and the pinned reconstruction source.
Internal spending limits and execution-authority metadata were removed from the three lock-file copies. Scientific settings remain in provenance/structured-output/, with source hashes and omitted fields recorded in PUBLICATION.json. Those projections are not executable authorization or byte-identical scientific locks. Model weights, private credentials, private human-audit labels, raw WikiText passages and generated study outputs are outside this release.
Loading and exact replay
from datasets import load_dataset
# Resolve public-v3 once and record its full Hub commit SHA in your experiment.
revision = "public-v3"
prose = load_dataset("euisuh/WONI-CrossBench", "cross-scheme", revision=revision)
structured = load_dataset("euisuh/WONI-CrossBench", "structured-output", revision=revision)
Use the full resolved commit SHA in the paper's artifact record. Verify files
against MANIFEST.json, which hashes every released file except
itself. o60871 contains a marker in its schema and rendered prompts; do not
silently score that modified prompt as an exact replication. To reconstruct the
three affected files in a new, private directory:
python hydrate.py /path/to/downloaded/snapshot /path/to/private-reconstruction
The hydrator uses datasets (verified with 3.2.0), obtains only the
pinned Github_hard source schema, verifies the withheld-value hash, and checks
all three reconstructed files against their original SHA-256 values. It does
not execute schema contents or contact the embedded endpoint. Reconstructed
material remains excluded from public uploads. Original scientific inputs and
public bytes have separate hashes in PUBLICATION.json.
Fields and evaluation limits
Prose prompt retains its trailing newline. Structured prompt and messages
use prompt_clarity_v1; legacy_prompt retains the earlier rendering. The
prompt-only and grammar-constrained conditions share a message; the latter
applies its grammar at decoding. Templates:
prompt-clarity-templates.json.
schema_json, reference_json and ground_truth_json encode heterogeneous
objects as JSON strings. Controlled schema-rendering prompts intentionally
include the target record. BFCL answer fields are evaluation material and are
not inserted into BFCL prompts; do not append them when evaluating.
Public experimental watermark keys and random seeds are reproducibility
parameters, not deployment secrets. Sources: the bundled scientific settings,
factorial inputs, and task files.
Sources and reuse
- Prose: Jens Abrahamsson's MIT-licensed text-watermark-laboratory at 9890649. Full original notice: prose-prompts-MIT.txt.
- 150 schema-rendering tasks: JSONSchemaBench at 5bd0f46. Its pinned card declares MIT. Target values and instructions are benchmark transformations. That declaration is upstream evidence, not an independent rights audit of every source project.
- 300 function-calling tasks: BFCL at 61fc060. Pinned card, Apache-2.0 text. Function schemas and output instructions are normalized; source IDs and reference semantics are retained.
See NOTICE.md. There is no blanket MIT grant over this mixed-source release. Public availability does not establish unrestricted redistribution.
Related WONI resources
| Resource | Role |
|---|---|
| WONI-correctness | Existing code-candidate evaluation labels and audit summaries; generated solution strings withheld |
| WONI-robustness | Exploratory token-survival summaries and human-null detector scores; WikiText token passages withheld |
| WONI-crossScheme | Existing factorial reader scores and sensitivity analyses; distinct from this input release |
The detector paper's generated-output evidence, replay code and WikiText reconstruction instructions are in the conformance artifact v1.2.0. The companion repositories do not establish robustness or correctness beyond their stated experiments. This named resource is not an anonymous review artifact.
September 24 exploratory prompt addendum
positive-control-2026-09-24/ adds 32
agent-authored tasks and 192 exact saved model/format prompt records for a
separate exploratory open-content control. It is not pooled with the primary
study. The public-v3 tag and all existing benchmark data remain unchanged;
select this commit (not public-v3) to obtain the addendum. No generated outputs
are released here. Sources and scope: the addendum's PROVENANCE.json and README.
MANIFEST.json records the additional files and the updated card hash.
Citation
@misc{jeong2026woni_crossbench,
author = {Euisuh Jeong},
title = {WONI-CrossBench: A Cross-Scheme Benchmark for Watermark Robustness, Detection Correctness, and Transfer},
year = {2026},
publisher = {Hugging Face},
howpublished = {\url{https://huggingface.co/datasets/euisuh/WONI-CrossBench}},
note = {Public input release 2026-09-23-public-v3; record the resolved commit SHA}
}
Also cite upstream datasets and the applicable study. For anonymous review, follow the venue's artifact rules; do not describe this author-linked URL as anonymous. The citation is to released inputs, not evidence of paper acceptance.
- Downloads last month
- 188