code · freeautoevalsdeterministicmirrors autoevals.Levenshtein

Levenshtein

Levenshtein returns 1 minus the normalised edit distance between the output and the expected string. It mirrors autoevals Levenshtein and is free.

how it works // frame by frame

What the evaluator does

Compute the character edit distance, divide by the longer length, subtract from 1.

frame 1/4 · the two inputs
output
calibrated probabilties
expected
calibrated probabilities
1.edit distance = 1 (insert 'i')
2.max length = 24
3.1 − 1/24
0.96score 0.96
runs in code · $0 · ~0 ms · deterministic
when to use it

Reach for it when

  • Near-exact string tasks: OCR, transcription, code identifiers.
  • A graded companion to exact_match.

Not the right tool when

  • Semantic tasks; a synonym is a large edit distance.

Watch out for

  • Length-sensitive: one wrong word in a long sentence scores high.
inputs and example

What to send

required fields
outputexpected
optional fields
none
in suites
CorrectnessEverything
result shape

score 0–1 computed in code, label, passed at an evaluator-specific threshold, no probabilities.

POST /api/v1/evaluateopen in console
{
  "evaluators": [
    "levenshtein"
  ],
  "output": "calibrated probabilties",
  "expected": "calibrated probabilities"
}
expected

One insertion over 24 characters: 0.96.

rolling it up

Aggregate with

One result per item is a fact; a dataset of them is a metric. These are the aggregations that fit this evaluator's output shape.

related evaluators
questions people ask
What is a good Levenshtein threshold?
jeval passes at 0.8 by default. For OCR-style tasks 0.95 is common; set it from your own data.