Gate analysis
Try: nahi / nhi / nahin, stretched letters avadlaaa, all-caps, or different spellings.
Decision
Analysis log
Dataset
Drop a labeled CSV here or click to browse
Columns needed:
text, label — leave empty to use the built-in sampleProgress
Flip rate by change type
Flip examples
| Type | Original | Variant | Answer changed |
|---|
No flips recorded.
What is FlipRate?
FlipRate measures how often an NLP model changes its answer when the same message is typed differently. People write "nahi", "nhi", "nahin" and "avadlaaa" for the same meaning. Many models answer differently for each variant. FlipRate calls this a flip.
Gate decisions
| Decision | Meaning |
|---|---|
| PASS | Stable across safe variants. |
| TRANSFORM | Unstable, but normalization fixes it. |
| REVIEW | Unstable and normalization does not help. |
API endpoints
POST /jobs/analyze— start gate analysis, get job_idPOST /jobs/test— start robustness testGET /jobs/{id}/stream— Server-Sent Events streamGET /health— {"status":"ok"}GET /ready— 503 until model loadsGET /metrics— Prometheus counters