scenario_id string | scenario_text string | claim string | label int64 |
|---|---|---|---|
train_001 | The patient has a temperature of 38.6°C and reports chills. | The patient has a fever. | 1 |
train_002 | The project missed two deadlines and three team members reported unclear ownership. | The project failed because the team is lazy. | 0 |
train_003 | The server response time increased from 180ms to 920ms after the latest deployment. | The system slowed after the latest deployment. | 1 |
train_004 | A customer cancelled their subscription after submitting two unresolved support tickets. | The customer hated the product. | 0 |
train_005 | Inventory dropped by 40% while weekly orders stayed unchanged. | Stock levels declined faster than expected demand. | 1 |
train_006 | The student scored 92% on the exam after attending all revision sessions. | The revision sessions caused the high score. | 0 |
train_007 | The sensor recorded rising vibration levels over three consecutive hours. | Vibration increased over time. | 1 |
train_008 | The employee was silent during the meeting and left early. | The employee disagreed with the proposal. | 0 |
train_009 | The app crash rate rose from 1.2% to 6.8% after the new release. | The new release coincided with a higher crash rate. | 1 |
train_010 | The tenant reported damp patches after heavy rain and photos show dark staining near the ceiling. | There is visible staining consistent with reported damp. | 1 |
train_011 | The company reduced hiring after revenue growth slowed for two quarters. | Leadership panicked. | 0 |
train_012 | Blood pressure fell from 128/82 to 96/60 after medication was given. | Blood pressure decreased after medication. | 1 |
train_013 | Three users complained about confusing navigation in the checkout flow. | All users find the checkout unusable. | 0 |
train_014 | The machine stopped twice during peak load and restarted after cooling. | The machine had stoppages during peak load. | 1 |
train_015 | The council delayed approval after requesting more environmental data. | The council opposes the project. | 0 |
train_016 | The child coughed throughout the night and had a runny nose. | The child showed respiratory symptoms. | 1 |
train_017 | Sales rose after the new advert campaign launched. | The advert campaign caused the sales increase. | 0 |
train_018 | Two suppliers missed delivery windows during the same week. | Supplier delivery reliability decreased that week. | 1 |
train_019 | The model gave different answers to the same prompt across three runs. | The model showed output inconsistency. | 1 |
train_020 | The manager did not reply for four days. | The manager is avoiding responsibility. | 0 |
What this dataset does
This dataset tests whether a model can separate direct observation from unsupported inference.
The task is simple:
Given a scenario and a claim, predict whether the claim stays within the evidence.
Core stability idea
Many reasoning failures begin when inference is treated as observation.
This dataset targets that failure mode.
A valid claim may describe what is directly shown in the scenario.
An invalid claim adds unsupported motive, cause, intent, severity, or certainty.
Prediction target
Binary label:
- 1 = the claim is supported by the scenario
- 0 = the claim goes beyond the scenario
Row structure
Each row contains:
- scenario_id
- scenario_text
- claim
- label
Files
- data/train.csv
- data/test.csv
- scorer.py
- README.md
Evaluation
Create a predictions CSV with:
scenario_id,prediction
test_001,1
test_002,0
Run:
python scorer.py --predictions predictions.csv --truth data/test.csv
The scorer reports:
accuracy
precision
recall
f1
confusion matrix
Structural Note
This dataset is intentionally small.
Its purpose is not scale.
Its purpose is to test whether a model can preserve epistemic discipline under simple conditions.
The hidden value is in the distinction between observation, inference, assumption, and overclaiming.
License
MIT
- Downloads last month
- 9