File size: 1,241 Bytes
1446570
 
a5bbc9e
 
 
1446570
a5bbc9e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
---
license: mit
language:
- he
pretty_name: MILIM-Bench
---

# MILIM-Bench

MILIM-Bench is a small Hebrew benchmark for evaluating grapheme-to-phoneme
(G2P) systems. It contains Hebrew text examples grouped by linguistic category,
with target pronunciations for the words that should be checked.

The benchmark is distributed as `MILIM-Bench-v5.tsv` and is hosted on Hugging
Face at `renikud/MILIM-Bench`.

## Format

Each row contains:

- `Category`: the phenomenon being tested.
- `Text`: the Hebrew sentence or phrase.
- `Label`: expected pronunciation targets, written as `token_index=IPA`.
  Token indexes start from 0, and a row may contain one or more indexed words.

Example label:

```text
2=leχˈa
```

This means token 2 in the text should be pronounced `leχˈa`.

## Categories

The current file has 1,653 examples across 12 categories:

- acronyms
- colloquial
- foreign
- gender
- homographs
- ilspeech-v2-test
- minimal stress pairs
- names
- penultimate stress
- rare phonemes
- slang
- Regular Homographed

## Intended Use

Use this dataset to compare Hebrew G2P outputs against category-specific
expected pronunciations, especially for ambiguity, stress placement, names,
loanwords, slang, acronyms, and gendered forms.