Fix bug with labels of eurlex config of lex_glue dataset (#5048)

* Update lex_glue.py

Fix for a critical bug in the EURLEX dataset label list to make LexGLUE EURLEX results replicable. In LexGLUE (Chalkidis et al., 2022), the following is mentioned w.r.t. EUR-LEX: "It supports four different label granularities, comprising 21, 127, 567, 7390 EuroVoc concepts, respectively. We use the 100 most frequent concepts from level 2 [...]”. The current label list has all 127 labels, which leads to different (lower) results, as communicated by users.

* Update lex_glue.py

Update code formatting

* Update dataset card

* Update metadata JSON

Co-authored-by: Albert Villanova del Moral <[email protected]>

Author: iliaschalkidis

datasets/lex_glue/README.md

@@ -199,7 +199,7 @@ An example of 'train' looks as follows.
 ```json
 {
   "text": "COMMISSION REGULATION (EC) No 1629/96 of 13 August 1996 on an invitation to tender for the refund on export of wholly milled round grain rice to certain third countries ...",
-  "labels": [2, 42, 72, 76, 86]
+  "labels": [4, 20, 21, 35, 68]
 }
 ```