Hugging Face's logo

Datasets:
HimangY
/
Human-Multi-Reference-MT-Benchmark

Tasks:
Translation
Languages:
English
Hindi
Telugu
Size:
1K<n<10K
Tags:
general
License:
Dataset card Data Studio Files
xet
Community
Human-Multi-Reference-MT-Benchmark
File size: 2,567 Bytes 
f1a938f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
 
---
license: cc
task_categories:
- translation
language:
- en
- hi
- te
tags:
- general
pretty_name: ' Human-Multi-Reference-MT-Benchmark'
size_categories:
- 1K<n<10K
---
# Human Benchmark Version v3

Parallel sentence benchmarks for the following language pairs:

- English–Hindi
- English–Telugu
- Hindi–Telugu
  
This repository includes multiple domains, reference translations, and
balanced splits for development and evaluation.

## Authors

- Vandan Mujadia
- Dipti Misra Sharma

## Acknowledgment

Developed as part of Himangy, LTRC IIIT H.

## Source Data Layout

Raw data is organized by language pair and domain:

- `English-Hindi/`
  - `wiki-articles/`
  - `wiki-news/`
  - `news-on-air/`
- `English-Telugu/online/`
  - `wiki-articles/`
  - `wiki-news/`
  - `news-on-air/`

Each raw `.txt` file is tab-separated. The first three columns are:

1. Task or passage id (varies by file)
2. Source sentence (English)
3. Target sentence (Hindi or Telugu)

## Processed Outputs

All generated files are stored under `outputs/`.

### 1) Flat Parallel Pairs

Source–target TSVs (one sentence pair per line):

- `outputs/english-hindi.tsv`
- `outputs/english-telugu.tsv`
- `outputs/hindi-telugu.tsv`

For Hindi–Telugu, pairs are created by matching shared English sentences
across the English–Hindi and English–Telugu datasets.

### 2) Grouped References (per source)

One row per source sentence, followed by all references:

- `outputs/english-hindi.references.tsv`
- `outputs/english-telugu.references.tsv`
- `outputs/hindi-telugu.references.tsv`

Format:

```
source<TAB>reference1<TAB>reference2<TAB>reference3...
```

### 3) Domain-Specific Splits with References

Split files are created per language pair, per domain, and per reference
(ref1, ref2, ref3). Each split has an equal number of examples.

Location:

```
outputs/splits/
```

Filename pattern:

```
{pair}.{domain}.ref{1|2|3}.{dev|devtest|test}.tsv
```

Example:

```
english-hindi.wiki-news.ref1.dev.tsv
```

### 4) Grouped by Language and Reference

Domain-concatenated files grouped by language pair and reference:

```
outputs/grouped/{pair}/ref{1|2|3}/{dev|devtest|test}.tsv
```

Example:

```
outputs/grouped/english-hindi/ref1/dev.tsv
```

## Splitting Policy

For each language pair and domain, sources are shuffled and split into
`dev`, `devtest`, and `test` with equal sizes. Any remainder that would
make the splits uneven is dropped.

## Notes

- All files are UTF-8.
- Tabs are used as field separators.
- If a source has fewer than 3 references, ref2/ref3 may be empty.