Update README.md
Browse files
README.md
CHANGED
@@ -33,7 +33,69 @@ size_categories:
|
|
33 |
|
34 |
For a better dataset description, please visit: [LINK](https://klue-benchmark.com/) <br>
|
35 |
<br>
|
36 |
-
**This dataset was prepared by converting KLUENLI dataset** to use it for contrastive training (SimCSE).
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
37 |
|
38 |
**How to download**
|
39 |
|
|
|
33 |
|
34 |
For a better dataset description, please visit: [LINK](https://klue-benchmark.com/) <br>
|
35 |
<br>
|
36 |
+
**This dataset was prepared by converting KLUENLI dataset** to use it for contrastive training (SimCSE). The code used to prepare the data is given below:
|
37 |
+
|
38 |
+
```py
|
39 |
+
import pandas as pd
|
40 |
+
from datasets import load_dataset, concatenate_datasets, Dataset
|
41 |
+
from torch.utils.data import random_split
|
42 |
+
|
43 |
+
|
44 |
+
class PrepTriplets:
|
45 |
+
@staticmethod
|
46 |
+
def make_dataset():
|
47 |
+
train_dataset = load_dataset("klue", "nli", split="train")
|
48 |
+
val_dataset = load_dataset("klue", "nli", split="validation")
|
49 |
+
merged_dataset = concatenate_datasets([train_dataset, val_dataset])
|
50 |
+
|
51 |
+
triplets_dataset = PrepTriplets._get_triplets(merged_dataset)
|
52 |
+
|
53 |
+
# Split back into train and validation
|
54 |
+
train_size = int(0.9 * len(triplets_dataset))
|
55 |
+
val_size = len(triplets_dataset) - train_size
|
56 |
+
train_subset, val_subset = random_split(
|
57 |
+
triplets_dataset, [train_size, val_size]
|
58 |
+
)
|
59 |
+
|
60 |
+
# Convert Subset objects back to Dataset
|
61 |
+
train_dataset = triplets_dataset.select(train_subset.indices)
|
62 |
+
val_dataset = triplets_dataset.select(val_subset.indices)
|
63 |
+
|
64 |
+
return train_dataset, val_dataset
|
65 |
+
|
66 |
+
@staticmethod
|
67 |
+
def _get_triplets(dataset):
|
68 |
+
df = pd.DataFrame(dataset)
|
69 |
+
|
70 |
+
entailments = df[df["label"] == 0]
|
71 |
+
contradictions = df[df["label"] == 2]
|
72 |
+
|
73 |
+
triplets = []
|
74 |
+
|
75 |
+
for premise in df["premise"].unique():
|
76 |
+
entailment_hypothesis = entailments[entailments["premise"] == premise][
|
77 |
+
"hypothesis"
|
78 |
+
].tolist()
|
79 |
+
contradiction_hypothesis = contradictions[
|
80 |
+
contradictions["premise"] == premise
|
81 |
+
]["hypothesis"].tolist()
|
82 |
+
|
83 |
+
if entailment_hypothesis and contradiction_hypothesis:
|
84 |
+
triplets.append(
|
85 |
+
{
|
86 |
+
"premise": premise,
|
87 |
+
"entailment": entailment_hypothesis[0],
|
88 |
+
"contradiction": contradiction_hypothesis[0],
|
89 |
+
}
|
90 |
+
)
|
91 |
+
|
92 |
+
triplets_dataset = Dataset.from_pandas(pd.DataFrame(triplets))
|
93 |
+
|
94 |
+
return triplets_dataset
|
95 |
+
|
96 |
+
# Example usage:
|
97 |
+
# PrepTriplets.make_dataset()
|
98 |
+
```
|
99 |
|
100 |
**How to download**
|
101 |
|