ghomasHudson commited on
Commit
a88bfdd
1 Parent(s): 7a3fb90

Update muld.py

Browse files
Files changed (1) hide show
  1. muld.py +10 -8
muld.py CHANGED
@@ -1,4 +1,4 @@
1
- """MuLD: The Multitask Long Document Benchmark"""
2
 
3
  import json
4
  import os
@@ -38,18 +38,19 @@ The NarrativeQA Reading Comprehension Challenge Dataset consists of user-submitt
38
  publisher={MIT Press}
39
  }""",
40
  "urls": {
41
- datasets.Split.TRAIN: "https://drive.google.com/uc?export=download&id=1sUXIC6lmk9Khp2mnr9VZwQ-StDlHqTw1?confirm=t",
42
- datasets.Split.VALIDATION: "https://drive.google.com/uc?export=download&id=1xdXEhLHtcqOZh0FbPhY_dnvNMg2bALtm",
43
  datasets.Split.TEST: "https://drive.google.com/uc?export=download&id=1BPBXyfYWVGtOXVQv_hlqtvbT25rTQzGu",
44
  }
45
  },
46
 
47
  "HotpotQA": {
48
  "description": """\
49
- The HotpotQA dataset consists of questions from crowd workers which require information from multiple Wikipedia articles in order to answer, thus testing the ability for models to perform multi-hop question answering. The data is commonly presented as a list of paragraphs containing relevant information plus a setting where the addition of 'distractor paragraphs' fully test the ability of the model to comprehend which information is relevant to the question asked. To transform this into a long document, we expand each paragraph with its full Wikipedia page as well as adding additional distractor articles from similar topics (randomly chosen from links on the existing pages) in order to meet the 10,000 token minimum length requirement for this benchmark. These articles are shuffled and concatenated to form the model input.""",
 
50
  "urls": {
51
- datasets.Split.TRAIN: "https://drive.google.com/uc?export=download&id=1OlGRyCEL9JhwIQIKViaWIXCOB_pwj8xU?confirm=t",
52
- datasets.Split.VALIDATION: "https://drive.google.com/uc?export=download&id=1_Svtg6PycBpezDYJ78zcJqLa8Ohnk6Gq?confirm=t"
53
  }
54
  },
55
 
@@ -92,7 +93,8 @@ Style change detection is the task of identifying the points where the author ch
92
 
93
  "VLSP": {
94
  "description": """\
95
- We follow the process of the Scientific papers (Cohan et al.,2018) summarization dataset, extracting papers from the open-access preprint server Arxiv.org using both the arxiv short abstract and the one included in the document (where available) as the reference summaries. In contrast to Cohan et al. (2018), rather than removing very long documents, we explicitly include them - removing any document with less than 10,000 tokens.""",
 
96
  "urls": {
97
  datasets.Split.TEST: "https://drive.google.com/uc?export=download&id=1ljTZZV5MpD07my2Vn1SVT3eQPKMVlHU5"
98
  }
@@ -153,4 +155,4 @@ class Muld(datasets.GeneratorBasedBuilder):
153
  row = json.loads(line)
154
  if "metadata" not in row:
155
  row["metadata"] = ""
156
- yield idx, row
 
1
+ """The MuLD benchmark."""
2
 
3
  import json
4
  import os
 
38
  publisher={MIT Press}
39
  }""",
40
  "urls": {
41
+ datasets.Split.TRAIN: "https://drive.google.com/uc?export=download&confirm=yTib&id=1sUXIC6lmk9Khp2mnr9VZwQ-StDlHqTw1",
42
+ datasets.Split.VALIDATION: "https://drive.google.com/uc?export=downloadid=1xdXEhLHtcqOZh0FbPhY_dnvNMg2bALtm",
43
  datasets.Split.TEST: "https://drive.google.com/uc?export=download&id=1BPBXyfYWVGtOXVQv_hlqtvbT25rTQzGu",
44
  }
45
  },
46
 
47
  "HotpotQA": {
48
  "description": """\
49
+ The HotpotQA dataset consists of questions from crowd workers which require information from multiple Wikipedia articles in order to answer, thus testing the ability for models to perform multi-hop question answering. The data is commonly presented as a list of paragraphs containing relevant information plus a setting where the addition of 'distractor paragraphs' fully test the ability of the model to comprehend which information is relevant to the question asked. To transform this into a long document, we expand each paragraph with its full Wikipedia page as well as adding additional distractor articles
50
+ from similar topics (randomly chosen from links on the existing pages) in order to meet the 10,000 token minimum length requirement for this benchmark. These articles are shuffled and concatenated to form the model input.""",
51
  "urls": {
52
+ datasets.Split.TRAIN: "https://drive.google.com/uc?export=download&confirm=yTib&id=1OlGRyCEL9JhwIQIKViaWIXCOB_pwj8xU",
53
+ datasets.Split.VALIDATION: "https://drive.google.com/uc?export=download&confirm=yTib&id=1_Svtg6PycBpezDYJ78zcJqLa8Ohnk6Gq"
54
  }
55
  },
56
 
 
93
 
94
  "VLSP": {
95
  "description": """\
96
+ We follow the process of the Scientific papers (Cohan et al.,2018) summarization dataset, extracting papers from the open-access preprint server Arxiv.org using both the arxiv short abstract and the one included in the document (where available) as the reference summaries. In contrast to Cohan et al.
97
+ (2018), rather than removing very long documents, we explicitly include them - removing any document with less than 10,000 tokens.""",
98
  "urls": {
99
  datasets.Split.TEST: "https://drive.google.com/uc?export=download&id=1ljTZZV5MpD07my2Vn1SVT3eQPKMVlHU5"
100
  }
 
155
  row = json.loads(line)
156
  if "metadata" not in row:
157
  row["metadata"] = ""
158
+ yield idx, row