pietrolesci commited on
Commit
25a1645
1 Parent(s): 74f4918

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +19 -0
README.md CHANGED
@@ -10,6 +10,9 @@ The only differences are the following:
10
  - renaming columns with the following mapping `{"sentence1": "premise", "sentence2": "hypothesis", "gold_label": "label"}`
11
  - encoding labels with the following mapping `{"entailment": 0, "neutral": 1, "contradiction": 2}`
12
 
 
 
 
13
 
14
  ## Code to create the dataset
15
  ```python
@@ -41,4 +44,20 @@ for name, df_ in dd.items():
41
 
42
  dataset = DatasetDict(ds)
43
  dataset.push_to_hub("scitail", token="<token>")
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
44
  ```
 
10
  - renaming columns with the following mapping `{"sentence1": "premise", "sentence2": "hypothesis", "gold_label": "label"}`
11
  - encoding labels with the following mapping `{"entailment": 0, "neutral": 1, "contradiction": 2}`
12
 
13
+ Note that there are 10 overlapping instances (as found by merging on columns "label", "premise", and "hypothesis") between
14
+ `train` and `test` splits.
15
+
16
 
17
  ## Code to create the dataset
18
  ```python
 
44
 
45
  dataset = DatasetDict(ds)
46
  dataset.push_to_hub("scitail", token="<token>")
47
+
48
+ # check overlap between splits
49
+ from itertools import combinations
50
+ for i, j in combinations(dataset.keys(), 2):
51
+ print(
52
+ f"{i} - {j}: ",
53
+ pd.merge(
54
+ dataset[i].to_pandas(),
55
+ dataset[j].to_pandas(),
56
+ on=["label", "premise", "hypothesis"],
57
+ how="inner",
58
+ ).shape[0],
59
+ )
60
+ #> train - test: 10
61
+ #> train - validation: 0
62
+ #> test - validation: 0
63
  ```