id
stringlengths 7
7
| domain
stringclasses 8
values | source_text
stringlengths 17
48
| boundary_markers
stringclasses 10
values | tts_transcript_with_marks
stringlengths 17
48
| boundary_pressure
stringclasses 8
values | constraints
stringclasses 4
values | failure_modes_to_avoid
stringlengths 11
28
| target_behaviors
stringlengths 9
25
| gold_checklist
stringlengths 4
24
|
|---|---|---|---|---|---|---|---|---|---|
SBI-001
|
health_info
|
This may reduce pain.
|
may
|
This MAY reduce pain.
|
modal_loss
|
Under 40 words.
|
May -> will
|
Keep MAY emphasized.
|
1 MAY present; 2 no WILL
|
SBI-002
|
health_info
|
This might help some people.
|
might,some
|
This MIGHT help SOME people.
|
scope_gate_loss
|
Under 50 words.
|
Some -> all
|
Keep SOME and MIGHT.
|
1 SOME; 2 MIGHT
|
SBI-003
|
medical
|
Do not take with alcohol.
|
not
|
Do NOT take with alcohol.
|
negation_loss
|
Under 40 words.
|
Negation softened
|
Keep NOT emphasized.
|
1 NOT
|
SBI-004
|
medical
|
Call a doctor if symptoms worsen.
|
if
|
Call a doctor IF symptoms worsen.
|
conditional_drop
|
Under 60 words.
|
Dropping IF
|
Keep IF clause clear.
|
1 IF
|
SBI-005
|
medical
|
Only use for children over 12.
|
only,over 12
|
ONLY use for children over 12.
|
scope_gate_loss
|
Under 60 words.
|
Only removed
|
Keep ONLY and 12.
|
1 ONLY; 2 12
|
SBI-006
|
cardiology
|
Absolute risk drops from 2% to 1%.
|
2%,1%
|
Absolute risk drops from 2% to 1%.
|
numeric_blur
|
Under 60 words.
|
Rounding or dropping numbers
|
Keep both numbers intact.
|
1 2%; 2 1%
|
SBI-007
|
oncology
|
Median survival increased by 2.1 months.
|
2.1 months
|
Median survival increased by 2.1 months.
|
numeric_rounding
|
Under 60 words.
|
Rounding away meaning
|
Keep 2.1 months.
|
1 2.1 months
|
SBI-008
|
policy
|
You are not required to respond.
|
not
|
You are NOT required to respond.
|
negation_loss
|
Under 60 words.
|
Softened negation
|
Keep NOT.
|
1 NOT
|
SBI-009
|
product_info
|
Results may vary.
|
may
|
Results MAY vary.
|
modal_loss
|
Under 40 words.
|
May removed
|
Keep MAY.
|
1 MAY
|
SBI-010
|
medical
|
Do not stop suddenly unless your doctor says so.
|
not,unless
|
Do NOT stop suddenly UNLESS your doctor says so.
|
conditional_scope
|
Under 80 words.
|
Unless dropped
|
Keep NOT and UNLESS.
|
1 NOT; 2 UNLESS
|
SBI-011
|
legal_plain
|
You may cancel within 14 days.
|
may,14 days
|
You MAY cancel within 14 days.
|
right_softening
|
Under 60 words.
|
Right made vague
|
Keep MAY and 14 days.
|
1 MAY; 2 14
|
SBI-012
|
tech
|
This feature is available only on Pro plans.
|
only
|
This feature is available ONLY on Pro plans.
|
scope_gate_loss
|
Under 60 words.
|
Only removed
|
Keep ONLY.
|
1 ONLY
|
YAML Metadata
Warning:
The task_categories "speech" is not in the official list: text-classification, token-classification, table-question-answering, question-answering, zero-shot-classification, translation, summarization, feature-extraction, text-generation, fill-mask, sentence-similarity, text-to-speech, text-to-audio, automatic-speech-recognition, audio-to-audio, audio-classification, audio-text-to-text, voice-activity-detection, depth-estimation, image-classification, object-detection, image-segmentation, text-to-image, image-to-text, image-to-image, image-to-video, unconditional-image-generation, video-classification, reinforcement-learning, robotics, tabular-classification, tabular-regression, tabular-to-text, table-to-text, multiple-choice, text-ranking, text-retrieval, time-series-forecasting, text-to-video, image-text-to-text, image-text-to-image, image-text-to-video, visual-question-answering, document-question-answering, zero-shot-image-classification, graph-ml, mask-generation, zero-shot-object-detection, text-to-3d, image-to-3d, image-feature-extraction, video-text-to-text, keypoint-detection, visual-document-retrieval, any-to-any, video-to-video, other
What this dataset tests
Speech must preserve boundaries.
Negation matters.
Modality matters.
Conditions matter.
Numbers matter.
Why it exists
Voice systems can blur meaning.
May becomes will.
If disappears.
Only gets lost.
Numbers get rounded.
This set detects boundary loss.
Data format
Each row contains
- source_text
- boundary_markers
- tts_transcript_with_marks
- boundary_pressure
Inline marks stand in for audible emphasis.
What is scored
- boundary marker retention
- emphasis on critical markers
- numeric fidelity
Boundary pressures
- modal_loss
- scope_gate_loss
- negation_loss
- conditional_drop
- numeric_blur
- numeric_rounding
- conditional_scope
- right_softening
Questions you must answer
- Did modality stay modality
- Did negation stay negation
- Did conditions stay conditions
- Did numbers stay exact
Suggested prompt wrapper
System
You evaluate whether TTS preserves semantic boundaries using text cues.
User
Source Text
{source_text}
Boundary Markers
{boundary_markers}
TTS Transcript
{tts_transcript_with_marks}
Scoring
Use scorer.py.
It returns
- score from 0 to 1
- boundary coverage signals
Known failure signatures
- MAY -> WILL
- IF dropped
- ONLY removed
- numbers rounded
Citation
ClarusC64 dataset family
- Downloads last month
- 9