Release complete controlled WinoBias-UK dataset
Browse files- README.md +12 -10
- data/test.csv +0 -0
- data/validation.csv +0 -0
README.md
CHANGED
|
@@ -11,7 +11,7 @@ tags:
|
|
| 11 |
- bias-evaluation
|
| 12 |
- ukrainian
|
| 13 |
size_categories:
|
| 14 |
-
- n<
|
| 15 |
configs:
|
| 16 |
- config_name: default
|
| 17 |
data_files:
|
|
@@ -29,13 +29,14 @@ gender-neutral Ukrainian descriptions, while the evaluated pronoun remains gende
|
|
| 29 |
|
| 30 |
## Current release
|
| 31 |
|
| 32 |
-
This
|
| 33 |
|
| 34 |
-
- validation:
|
| 35 |
-
- test:
|
| 36 |
- each source item has one masculine-pronoun and one feminine-pronoun version.
|
| 37 |
|
| 38 |
-
|
|
|
|
| 39 |
|
| 40 |
## Fields
|
| 41 |
|
|
@@ -53,8 +54,9 @@ The full controlled construction is in progress. Future releases will expand the
|
|
| 53 |
|
| 54 |
The Ukrainian data was produced through machine-assisted translation and adaptation. The
|
| 55 |
occupation lexicon was reviewed and approved by the adaptation team. Structural validation checks
|
| 56 |
-
candidate balance, pro/anti pairing, required fields,
|
| 57 |
-
|
|
|
|
| 58 |
|
| 59 |
The released files contain Ukrainian evaluation data. Working prompts, API records, English source
|
| 60 |
sentences, and incomplete rows are excluded.
|
|
@@ -62,8 +64,9 @@ sentences, and incomplete rows are excluded.
|
|
| 62 |
## Limitations
|
| 63 |
|
| 64 |
WinoBias models binary grammatical gender and occupational stereotypes. Results should be reported
|
| 65 |
-
separately by type, condition, and pronoun gender.
|
| 66 |
-
and
|
|
|
|
| 67 |
|
| 68 |
## License
|
| 69 |
|
|
@@ -97,4 +100,3 @@ copyright and license notice are preserved in `LICENSE`.
|
|
| 97 |
note = {Ukrainian translation, adaptation, and cleaning of WinoBias}
|
| 98 |
}
|
| 99 |
```
|
| 100 |
-
|
|
|
|
| 11 |
- bias-evaluation
|
| 12 |
- ukrainian
|
| 13 |
size_categories:
|
| 14 |
+
- 1K<n<10K
|
| 15 |
configs:
|
| 16 |
- config_name: default
|
| 17 |
data_files:
|
|
|
|
| 29 |
|
| 30 |
## Current release
|
| 31 |
|
| 32 |
+
This release contains 1,578 source items and 3,156 evaluation rows:
|
| 33 |
|
| 34 |
+
- validation: 788 source items and 1,576 rows;
|
| 35 |
+
- test: 790 source items and 1,580 rows;
|
| 36 |
- each source item has one masculine-pronoun and one feminine-pronoun version.
|
| 37 |
|
| 38 |
+
Four original source pairs are omitted because `cleaner` and `janitor` map to the same controlled
|
| 39 |
+
Ukrainian occupation label and therefore cannot form two distinct answer candidates.
|
| 40 |
|
| 41 |
## Fields
|
| 42 |
|
|
|
|
| 54 |
|
| 55 |
The Ukrainian data was produced through machine-assisted translation and adaptation. The
|
| 56 |
occupation lexicon was reviewed and approved by the adaptation team. Structural validation checks
|
| 57 |
+
candidate balance, pro/anti pairing, pronoun forms, required fields, unique item keys, and exact
|
| 58 |
+
occupation-label placement. Every released pair also received item-level semantic review during
|
| 59 |
+
construction. Independent human linguistic validation remains future work.
|
| 60 |
|
| 61 |
The released files contain Ukrainian evaluation data. Working prompts, API records, English source
|
| 62 |
sentences, and incomplete rows are excluded.
|
|
|
|
| 64 |
## Limitations
|
| 65 |
|
| 66 |
WinoBias models binary grammatical gender and occupational stereotypes. Results should be reported
|
| 67 |
+
separately by type, condition, and pronoun gender. The controlled occupation descriptions prioritize
|
| 68 |
+
gender neutrality and comparability, so they can be less idiomatic than ordinary Ukrainian job
|
| 69 |
+
titles.
|
| 70 |
|
| 71 |
## License
|
| 72 |
|
|
|
|
| 100 |
note = {Ukrainian translation, adaptation, and cleaning of WinoBias}
|
| 101 |
}
|
| 102 |
```
|
|
|
data/test.csv
CHANGED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/validation.csv
CHANGED
|
The diff for this file is too large to render.
See raw diff
|
|
|