scenario_id string | scenario_text string | claim string | label int64 |
|---|---|---|---|
train_001 | Three calibrated sensors recorded the same pressure spike within 30 seconds. | The observation is reliable enough to use. | 1 |
train_002 | One uncalibrated sensor reported a pressure spike, while two calibrated sensors stayed normal. | The observation is reliable enough to use. | 0 |
train_003 | Two nurses independently recorded the same patient symptom during the same shift. | The observation is reliable enough to use. | 1 |
train_004 | A symptom was reported once by someone who did not examine the patient. | The observation is reliable enough to use. | 0 |
train_005 | The server logs, monitoring dashboard, and customer reports all show downtime at the same time. | The observation is reliable enough to use. | 1 |
train_006 | A dashboard shows downtime, but server logs show normal operation and no users reported issues. | The observation is reliable enough to use. | 0 |
train_007 | A warehouse count was checked twice by separate staff and matched both times. | The observation is reliable enough to use. | 1 |
train_008 | A warehouse count was estimated from memory after the stock system failed. | The observation is reliable enough to use. | 0 |
train_009 | The payment record appears in the bank statement, receipt, and invoice ledger. | The observation is reliable enough to use. | 1 |
train_010 | The payment record appears only in a manually edited spreadsheet with no timestamp. | The observation is reliable enough to use. | 0 |
train_011 | A model answer matches the source document and cites the correct section. | The observation is reliable enough to use. | 1 |
train_012 | A model answer cites a source that does not contain the stated claim. | The observation is reliable enough to use. | 0 |
train_013 | Humidity readings rose on two devices and visible damp staining expanded. | The observation is reliable enough to use. | 1 |
train_014 | Humidity was reported by one cheap device placed beside a kettle. | The observation is reliable enough to use. | 0 |
train_015 | A customer complaint includes screenshots, timestamps, and order number. | The observation is reliable enough to use. | 1 |
train_016 | A customer complaint says only 'it broke' with no product, time, or error details. | The observation is reliable enough to use. | 0 |
train_017 | A traffic counter was tested that morning and matched manual counts within 2%. | The observation is reliable enough to use. | 1 |
train_018 | A traffic count came from a damaged counter during heavy rain, with no validation. | The observation is reliable enough to use. | 0 |
train_019 | A medication record was signed by the nurse and matched the electronic prescribing system. | The observation is reliable enough to use. | 1 |
train_020 | A medication record was reconstructed later from memory after notes were lost. | The observation is reliable enough to use. | 0 |
What this dataset does
This dataset tests whether a model can judge whether an observation is reliable enough to reason from.
The task is simple:
Given a scenario and a reliability claim, predict whether the claim is supported.
Core stability idea
Reasoning fails when weak observations are treated as stable evidence.
This dataset targets that failure mode.
An observation is reliable when it is confirmed, timestamped, validated, independently repeated, or consistent across sources.
An observation is unreliable when it is isolated, contradicted, uncalibrated, vague, reconstructed from memory, or unsupported.
Prediction target
Binary label:
- 1 = the observation is reliable enough to use
- 0 = the observation is not reliable enough to use
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 to test whether a model can separate usable signal from weak observation.
The hidden value is in detecting source quality, independent confirmation, contradiction, calibration, traceability, and evidential fragility.
License
MIT
- Downloads last month
- -