Update dataset card
Browse files
README.md
CHANGED
|
@@ -26,7 +26,9 @@ configs:
|
|
| 26 |
path: data/dev_test-*
|
| 27 |
- split: test
|
| 28 |
path: data/test-*
|
|
|
|
| 29 |
dataset_info:
|
|
|
|
| 30 |
features:
|
| 31 |
- name: id
|
| 32 |
dtype: string
|
|
@@ -40,19 +42,13 @@ dataset_info:
|
|
| 40 |
sequence: string
|
| 41 |
splits:
|
| 42 |
- name: train
|
| 43 |
-
num_bytes: 274812860.5
|
| 44 |
num_examples: 3500
|
| 45 |
- name: dev
|
| 46 |
-
num_bytes: 40130069.0
|
| 47 |
num_examples: 500
|
| 48 |
- name: dev_test
|
| 49 |
-
num_bytes: 38926730.0
|
| 50 |
num_examples: 500
|
| 51 |
- name: test
|
| 52 |
-
num_bytes: 38669448.0
|
| 53 |
num_examples: 500
|
| 54 |
-
download_size: 386300009
|
| 55 |
-
dataset_size: 392539107.5
|
| 56 |
---
|
| 57 |
|
| 58 |
# ArGuard – Track A: Arabic Hateful Memes
|
|
@@ -85,12 +81,13 @@ Given a meme (image + Arabic text):
|
|
| 85 |
|--------------|---------|--------------|-----------------------------------------|---------------------------------|
|
| 86 |
| `train` | 3,500 | full | single-annotated bulk | development phase |
|
| 87 |
| `dev` | 500 | full | single-annotated bulk | development phase |
|
| 88 |
-
| `dev_test` | 500 |
|
| 89 |
| `test` | 500 | **dropped** | **triple-annotated gold** (calibration) | final-evaluation phase (leaderboard) |
|
| 90 |
|
| 91 |
-
- `dev_test` was the **leaderboard set** for the development phase
|
| 92 |
-
is now closed
|
| 93 |
-
`
|
|
|
|
| 94 |
- `test` is the **blind test set** for the final-evaluation phase and is
|
| 95 |
the file you submit predictions for. It is released the same way as
|
| 96 |
`dev_test` — images and OCR text only, labels stripped. All 500 records
|
|
@@ -104,7 +101,7 @@ Given a meme (image + Arabic text):
|
|
| 104 |
| train | 1,324 | 2,176 | 37.8% |
|
| 105 |
| dev | 189 | 311 | 37.8% |
|
| 106 |
| dev_test | 189 | 311 | 37.8% |
|
| 107 |
-
| **Total**
|
| 108 |
|
| 109 |
The label distribution of the blind `test` split is withheld during the
|
| 110 |
final-evaluation phase.
|
|
@@ -137,8 +134,8 @@ than the meme counts.
|
|
| 137 |
"id": "f9a8…b1.jpg", # str – original image filename, unique
|
| 138 |
"image": <PIL.Image.Image>, # embedded bytes, decoded on access
|
| 139 |
"text": "…", # str – OCR-extracted Arabic meme text
|
| 140 |
-
"label": "Hateful" | "Not Hateful" | None, # None on
|
| 141 |
-
"fine_grained_label": [...], # list[str] – empty on
|
| 142 |
}
|
| 143 |
```
|
| 144 |
|
|
@@ -154,13 +151,13 @@ train_ex = ds["train"][0]
|
|
| 154 |
train_ex["image"].show()
|
| 155 |
print(train_ex["text"], train_ex["label"], train_ex["fine_grained_label"])
|
| 156 |
|
| 157 |
-
#
|
| 158 |
print(ds["test"][0]["label"]) # -> None
|
| 159 |
```
|
| 160 |
|
| 161 |
The blind `test` split is the submission target for the final-evaluation
|
| 162 |
-
phase. Train on `train` + `dev`, predict on `test`, and
|
| 163 |
-
CodaBench (links below).
|
| 164 |
|
| 165 |
## Shared-task resources
|
| 166 |
|
|
|
|
| 26 |
path: data/dev_test-*
|
| 27 |
- split: test
|
| 28 |
path: data/test-*
|
| 29 |
+
default: true
|
| 30 |
dataset_info:
|
| 31 |
+
config_name: default
|
| 32 |
features:
|
| 33 |
- name: id
|
| 34 |
dtype: string
|
|
|
|
| 42 |
sequence: string
|
| 43 |
splits:
|
| 44 |
- name: train
|
|
|
|
| 45 |
num_examples: 3500
|
| 46 |
- name: dev
|
|
|
|
| 47 |
num_examples: 500
|
| 48 |
- name: dev_test
|
|
|
|
| 49 |
num_examples: 500
|
| 50 |
- name: test
|
|
|
|
| 51 |
num_examples: 500
|
|
|
|
|
|
|
| 52 |
---
|
| 53 |
|
| 54 |
# ArGuard – Track A: Arabic Hateful Memes
|
|
|
|
| 81 |
|--------------|---------|--------------|-----------------------------------------|---------------------------------|
|
| 82 |
| `train` | 3,500 | full | single-annotated bulk | development phase |
|
| 83 |
| `dev` | 500 | full | single-annotated bulk | development phase |
|
| 84 |
+
| `dev_test` | 500 | full | single-annotated test sample | development phase (leaderboard); **labels released** |
|
| 85 |
| `test` | 500 | **dropped** | **triple-annotated gold** (calibration) | final-evaluation phase (leaderboard) |
|
| 86 |
|
| 87 |
+
- `dev_test` was the **leaderboard set** for the development phase. That
|
| 88 |
+
phase is now closed and its gold labels have been **released**, so
|
| 89 |
+
`dev_test` is usable as additional labelled data. It is disjoint from
|
| 90 |
+
`test`.
|
| 91 |
- `test` is the **blind test set** for the final-evaluation phase and is
|
| 92 |
the file you submit predictions for. It is released the same way as
|
| 93 |
`dev_test` — images and OCR text only, labels stripped. All 500 records
|
|
|
|
| 101 |
| train | 1,324 | 2,176 | 37.8% |
|
| 102 |
| dev | 189 | 311 | 37.8% |
|
| 103 |
| dev_test | 189 | 311 | 37.8% |
|
| 104 |
+
| **Total (labelled)** | **1,702** | **2,798** | 37.8% |
|
| 105 |
|
| 106 |
The label distribution of the blind `test` split is withheld during the
|
| 107 |
final-evaluation phase.
|
|
|
|
| 134 |
"id": "f9a8…b1.jpg", # str – original image filename, unique
|
| 135 |
"image": <PIL.Image.Image>, # embedded bytes, decoded on access
|
| 136 |
"text": "…", # str – OCR-extracted Arabic meme text
|
| 137 |
+
"label": "Hateful" | "Not Hateful" | None, # None on the blind `test` split
|
| 138 |
+
"fine_grained_label": [...], # list[str] – empty on the blind `test` split
|
| 139 |
}
|
| 140 |
```
|
| 141 |
|
|
|
|
| 151 |
train_ex["image"].show()
|
| 152 |
print(train_ex["text"], train_ex["label"], train_ex["fine_grained_label"])
|
| 153 |
|
| 154 |
+
# `test` is the only unlabelled split — used only to produce submissions
|
| 155 |
print(ds["test"][0]["label"]) # -> None
|
| 156 |
```
|
| 157 |
|
| 158 |
The blind `test` split is the submission target for the final-evaluation
|
| 159 |
+
phase. Train on `train` + `dev` + `dev_test`, predict on `test`, and
|
| 160 |
+
upload to CodaBench (links below).
|
| 161 |
|
| 162 |
## Shared-task resources
|
| 163 |
|