Changes

              
    

          
          
        
        

            
f
1
{
f
1
{
            
2
  "access_rights": "",
2
  "access_rights": "",
            
3
  "author": "Mor Geva",
3
  "author": "Mor Geva",
            
4
  "author_email": "",
4
  "author_email": "",
            
5
  "citation": [
5
  "citation": [
            
6
    "https://doi.org/10.48550/arXiv.2406.13929",
6
    "https://doi.org/10.48550/arXiv.2406.13929",
            
7
    "https://doi.org/10.48550/arXiv.2212.09656",
7
    "https://doi.org/10.48550/arXiv.2212.09656",
            
8
    "https://doi.org/10.48550/arXiv.2403.01390",
8
    "https://doi.org/10.48550/arXiv.2403.01390",
            
9
    "https://doi.org/10.48550/arXiv.2402.10612",
9
    "https://doi.org/10.48550/arXiv.2402.10612",
            
10
    "https://doi.org/10.48550/arXiv.2402.18678",
10
    "https://doi.org/10.48550/arXiv.2402.18678",
            
11
    "https://doi.org/10.48550/arXiv.2309.13075"
11
    "https://doi.org/10.48550/arXiv.2309.13075"
            
12
  ],
12
  ],
            
13
  "creator_user_id": "17755db4-395a-4b3b-ac09-e8e3484ca700",
13
  "creator_user_id": "17755db4-395a-4b3b-ac09-e8e3484ca700",
            
14
  "defined_in": "https://doi.org/10.48550/arXiv.2302.12246",
14
  "defined_in": "https://doi.org/10.48550/arXiv.2302.12246",
            
15
  "doi": "10.57702/nkmvf3lj",
15
  "doi": "10.57702/nkmvf3lj",
            
n
16
  "doi_date_published": null,
n
16
  "doi_date_published": "2025-01-02",
            
17
  "doi_publisher": "TIB",
17
  "doi_publisher": "TIB",
            
n
18
  "doi_status": false,
n
18
  "doi_status": true,
            
19
  "domain": "https://service.tib.eu/ldmservice",
19
  "domain": "https://service.tib.eu/ldmservice",
            
20
  "extra_authors": [
20
  "extra_authors": [
            
21
    {
21
    {
            
22
      "extra_author": "Daniel Khashabi",
22
      "extra_author": "Daniel Khashabi",
            
23
      "orcid": ""
23
      "orcid": ""
            
24
    },
24
    },
            
25
    {
25
    {
            
26
      "extra_author": "Elad Segal",
26
      "extra_author": "Elad Segal",
            
27
      "orcid": ""
27
      "orcid": ""
            
28
    },
28
    },
            
29
    {
29
    {
            
30
      "extra_author": "Tushar Khot",
30
      "extra_author": "Tushar Khot",
            
31
      "orcid": ""
31
      "orcid": ""
            
32
    },
32
    },
            
33
    {
33
    {
            
34
      "extra_author": "Dan Roth",
34
      "extra_author": "Dan Roth",
            
35
      "orcid": ""
35
      "orcid": ""
            
36
    },
36
    },
            
37
    {
37
    {
            
38
      "extra_author": "Jonathan Berant",
38
      "extra_author": "Jonathan Berant",
            
39
      "orcid": ""
39
      "orcid": ""
            
40
    }
40
    }
            
41
  ],
41
  ],
            
42
  "groups": [
42
  "groups": [
            
43
    {
43
    {
            
44
      "description": "",
44
      "description": "",
            
45
      "display_name": "Commonsense Reasoning",
45
      "display_name": "Commonsense Reasoning",
            
46
      "id": "3be75146-9a95-41b8-9005-5774447957de",
46
      "id": "3be75146-9a95-41b8-9005-5774447957de",
            
47
      "image_display_url": "",
47
      "image_display_url": "",
            
48
      "name": "commonsense-reasoning",
48
      "name": "commonsense-reasoning",
            
49
      "title": "Commonsense Reasoning"
49
      "title": "Commonsense Reasoning"
            
50
    },
50
    },
            
51
    {
51
    {
            
52
      "description": "",
52
      "description": "",
            
53
      "display_name": "Multi-Hop",
53
      "display_name": "Multi-Hop",
            
54
      "id": "e0a73f25-d456-4171-907b-149dcfd9c52b",
54
      "id": "e0a73f25-d456-4171-907b-149dcfd9c52b",
            
55
      "image_display_url": "",
55
      "image_display_url": "",
            
56
      "name": "multi-hop",
56
      "name": "multi-hop",
            
57
      "title": "Multi-Hop"
57
      "title": "Multi-Hop"
            
58
    },
58
    },
            
59
    {
59
    {
            
60
      "description": "",
60
      "description": "",
            
61
      "display_name": "Question Answering",
61
      "display_name": "Question Answering",
            
62
      "id": "01e9bd6b-910d-421d-bb31-799604023211",
62
      "id": "01e9bd6b-910d-421d-bb31-799604023211",
            
63
      "image_display_url": "",
63
      "image_display_url": "",
            
64
      "name": "question-answering",
64
      "name": "question-answering",
            
65
      "title": "Question Answering"
65
      "title": "Question Answering"
            
66
    }
66
    }
            
67
  ],
67
  ],
            
68
  "id": "177dcac4-5208-4a79-8b3a-a555dc870013",
68
  "id": "177dcac4-5208-4a79-8b3a-a555dc870013",
            
69
  "isopen": false,
69
  "isopen": false,
            
70
  "landing_page": 
70
  "landing_page": 
            
71
"https://huggingface.co/datasets/metaeval/strategy-qa",
71
"https://huggingface.co/datasets/metaeval/strategy-qa",
            
72
  "license_title": null,
72
  "license_title": null,
            
73
  "link_orkg": "",
73
  "link_orkg": "",
            
74
  "metadata_created": "2025-01-02T21:31:17.177051",
74
  "metadata_created": "2025-01-02T21:31:17.177051",
            
n
75
  "metadata_modified": "2025-01-02T21:31:17.177059",
n
75
  "metadata_modified": "2025-01-02T21:31:17.745890",
            
76
  "name": "strategyqa",
76
  "name": "strategyqa",
            
77
  "notes": "The StrategyQA dataset is used to evaluate the ability of 
77
  "notes": "The StrategyQA dataset is used to evaluate the ability of 
            
78
LLMs in generating accurate answers to multi-step reasoning 
78
LLMs in generating accurate answers to multi-step reasoning 
            
79
questions.",
79
questions.",
            
n
80
  "num_resources": 0,
n
80
  "num_resources": 1,
            
81
  "num_tags": 16,
81
  "num_tags": 16,
            
82
  "organization": {
82
  "organization": {
            
83
    "approval_status": "approved",
83
    "approval_status": "approved",
            
84
    "created": "2024-11-25T12:11:38.292601",
84
    "created": "2024-11-25T12:11:38.292601",
            
85
    "description": "",
85
    "description": "",
            
86
    "id": "079d46db-32df-4b48-91f3-0a8bc8f69559",
86
    "id": "079d46db-32df-4b48-91f3-0a8bc8f69559",
            
87
    "image_url": "",
87
    "image_url": "",
            
88
    "is_organization": true,
88
    "is_organization": true,
            
89
    "name": "no-organization",
89
    "name": "no-organization",
            
90
    "state": "active",
90
    "state": "active",
            
91
    "title": "No Organization",
91
    "title": "No Organization",
            
92
    "type": "organization"
92
    "type": "organization"
            
93
  },
93
  },
            
94
  "owner_org": "079d46db-32df-4b48-91f3-0a8bc8f69559",
94
  "owner_org": "079d46db-32df-4b48-91f3-0a8bc8f69559",
            
95
  "private": false,
95
  "private": false,
            
96
  "relationships_as_object": [],
96
  "relationships_as_object": [],
            
97
  "relationships_as_subject": [],
97
  "relationships_as_subject": [],
            
t
98
  "resources": [],
t
98
  "resources": [
            
99
    {
            
100
      "cache_last_updated": null,
            
101
      "cache_url": null,
            
102
      "created": "2025-01-02T21:52:27",
            
103
      "data": [
            
104
        "dcterms:title",
            
105
        "dcterms:accessRights",
            
106
        "dcterms:creator",
            
107
        "dcterms:description",
            
108
        "dcterms:issued",
            
109
        "dcterms:language",
            
110
        "dcterms:identifier",
            
111
        "dcat:theme",
            
112
        "dcterms:type",
            
113
        "dcat:keyword",
            
114
        "dcat:landingPage",
            
115
        "dcterms:hasVersion",
            
116
        "dcterms:format",
            
117
        "mls:task",
            
118
        "datacite:isDescribedBy"
            
119
      ],
            
120
      "description": "The json representation of the dataset with its 
            
121
distributions based on DCAT.",
            
122
      "format": "JSON",
            
123
      "hash": "",
            
124
      "id": "68d73a19-2eff-4ec9-85a0-6b1bf3eb60da",
            
125
      "last_modified": "2025-01-02T21:31:17.736655",
            
126
      "metadata_modified": "2025-01-02T21:31:17.748725",
            
127
      "mimetype": "application/json",
            
128
      "mimetype_inner": null,
            
129
      "name": "Original Metadata",
            
130
      "package_id": "177dcac4-5208-4a79-8b3a-a555dc870013",
            
131
      "position": 0,
            
132
      "resource_type": null,
            
133
      "size": 1363,
            
134
      "state": "active",
            
135
      "url": 
            
136
resource/68d73a19-2eff-4ec9-85a0-6b1bf3eb60da/download/metadata.json",
            
137
      "url_type": "upload"
            
138
    }
            
139
  ],
            
99
  "services_used_list": "",
140
  "services_used_list": "",
            
100
  "state": "active",
141
  "state": "active",
            
101
  "tags": [
142
  "tags": [
            
102
    {
143
    {
            
103
      "display_name": "QA dataset",
144
      "display_name": "QA dataset",
            
104
      "id": "d3e45412-cfcd-4a0e-b91e-4eb408d1075b",
145
      "id": "d3e45412-cfcd-4a0e-b91e-4eb408d1075b",
            
105
      "name": "QA dataset",
146
      "name": "QA dataset",
            
106
      "state": "active",
147
      "state": "active",
            
107
      "vocabulary_id": null
148
      "vocabulary_id": null
            
108
    },
149
    },
            
109
    {
150
    {
            
110
      "display_name": "StrategyQA",
151
      "display_name": "StrategyQA",
            
111
      "id": "cb0be26f-de55-4ed3-8166-3d1e58b3f424",
152
      "id": "cb0be26f-de55-4ed3-8166-3d1e58b3f424",
            
112
      "name": "StrategyQA",
153
      "name": "StrategyQA",
            
113
      "state": "active",
154
      "state": "active",
            
114
      "vocabulary_id": null
155
      "vocabulary_id": null
            
115
    },
156
    },
            
116
    {
157
    {
            
117
      "display_name": "Wikipedia articles",
158
      "display_name": "Wikipedia articles",
            
118
      "id": "0b7378d3-9081-431f-9985-df07ef768343",
159
      "id": "0b7378d3-9081-431f-9985-df07ef768343",
            
119
      "name": "Wikipedia articles",
160
      "name": "Wikipedia articles",
            
120
      "state": "active",
161
      "state": "active",
            
121
      "vocabulary_id": null
162
      "vocabulary_id": null
            
122
    },
163
    },
            
123
    {
164
    {
            
124
      "display_name": "cognitive tasks",
165
      "display_name": "cognitive tasks",
            
125
      "id": "3e682194-c52f-465a-91c8-029cfc4825c5",
166
      "id": "3e682194-c52f-465a-91c8-029cfc4825c5",
            
126
      "name": "cognitive tasks",
167
      "name": "cognitive tasks",
            
127
      "state": "active",
168
      "state": "active",
            
128
      "vocabulary_id": null
169
      "vocabulary_id": null
            
129
    },
170
    },
            
130
    {
171
    {
            
131
      "display_name": "commonsense reasoning",
172
      "display_name": "commonsense reasoning",
            
132
      "id": "52066668-35e4-47a9-9cf9-3c8262f60dbb",
173
      "id": "52066668-35e4-47a9-9cf9-3c8262f60dbb",
            
133
      "name": "commonsense reasoning",
174
      "name": "commonsense reasoning",
            
134
      "state": "active",
175
      "state": "active",
            
135
      "vocabulary_id": null
176
      "vocabulary_id": null
            
136
    },
177
    },
            
137
    {
178
    {
            
138
      "display_name": "context-based question answering",
179
      "display_name": "context-based question answering",
            
139
      "id": "3913defa-f016-4c7d-8fa9-c6f87bc01ac6",
180
      "id": "3913defa-f016-4c7d-8fa9-c6f87bc01ac6",
            
140
      "name": "context-based question answering",
181
      "name": "context-based question answering",
            
141
      "state": "active",
182
      "state": "active",
            
142
      "vocabulary_id": null
183
      "vocabulary_id": null
            
143
    },
184
    },
            
144
    {
185
    {
            
145
      "display_name": "human-annotated",
186
      "display_name": "human-annotated",
            
146
      "id": "5503021d-9375-4a4d-b16c-6d670575ee98",
187
      "id": "5503021d-9375-4a4d-b16c-6d670575ee98",
            
147
      "name": "human-annotated",
188
      "name": "human-annotated",
            
148
      "state": "active",
189
      "state": "active",
            
149
      "vocabulary_id": null
190
      "vocabulary_id": null
            
150
    },
191
    },
            
151
    {
192
    {
            
152
      "display_name": "model-generated",
193
      "display_name": "model-generated",
            
153
      "id": "c00b58ec-556d-4df9-9c1f-e35deaf32abd",
194
      "id": "c00b58ec-556d-4df9-9c1f-e35deaf32abd",
            
154
      "name": "model-generated",
195
      "name": "model-generated",
            
155
      "state": "active",
196
      "state": "active",
            
156
      "vocabulary_id": null
197
      "vocabulary_id": null
            
157
    },
198
    },
            
158
    {
199
    {
            
159
      "display_name": "multi-hop",
200
      "display_name": "multi-hop",
            
160
      "id": "3efd04a4-c30b-41c8-9799-45bb552de0f4",
201
      "id": "3efd04a4-c30b-41c8-9799-45bb552de0f4",
            
161
      "name": "multi-hop",
202
      "name": "multi-hop",
            
162
      "state": "active",
203
      "state": "active",
            
163
      "vocabulary_id": null
204
      "vocabulary_id": null
            
164
    },
205
    },
            
165
    {
206
    {
            
166
      "display_name": "multi-step reasoning",
207
      "display_name": "multi-step reasoning",
            
167
      "id": "da607e31-92fd-474b-b0f2-326f634ca6d0",
208
      "id": "da607e31-92fd-474b-b0f2-326f634ca6d0",
            
168
      "name": "multi-step reasoning",
209
      "name": "multi-step reasoning",
            
169
      "state": "active",
210
      "state": "active",
            
170
      "vocabulary_id": null
211
      "vocabulary_id": null
            
171
    },
212
    },
            
172
    {
213
    {
            
173
      "display_name": "open-domain questions",
214
      "display_name": "open-domain questions",
            
174
      "id": "f78262a8-cbf5-4c1b-8d40-b3b0bed6ef29",
215
      "id": "f78262a8-cbf5-4c1b-8d40-b3b0bed6ef29",
            
175
      "name": "open-domain questions",
216
      "name": "open-domain questions",
            
176
      "state": "active",
217
      "state": "active",
            
177
      "vocabulary_id": null
218
      "vocabulary_id": null
            
178
    },
219
    },
            
179
    {
220
    {
            
180
      "display_name": "question answering",
221
      "display_name": "question answering",
            
181
      "id": "d4e35d0b-ca07-4d05-bb27-50b2f30f2f00",
222
      "id": "d4e35d0b-ca07-4d05-bb27-50b2f30f2f00",
            
182
      "name": "question answering",
223
      "name": "question answering",
            
183
      "state": "active",
224
      "state": "active",
            
184
      "vocabulary_id": null
225
      "vocabulary_id": null
            
185
    },
226
    },
            
186
    {
227
    {
            
187
      "display_name": "reasoning steps",
228
      "display_name": "reasoning steps",
            
188
      "id": "c0b8f360-d155-4ba9-b994-aead31d27930",
229
      "id": "c0b8f360-d155-4ba9-b994-aead31d27930",
            
189
      "name": "reasoning steps",
230
      "name": "reasoning steps",
            
190
      "state": "active",
231
      "state": "active",
            
191
      "vocabulary_id": null
232
      "vocabulary_id": null
            
192
    },
233
    },
            
193
    {
234
    {
            
194
      "display_name": "strategyqa",
235
      "display_name": "strategyqa",
            
195
      "id": "a2530fe0-af46-4bf2-8387-309181bb08bc",
236
      "id": "a2530fe0-af46-4bf2-8387-309181bb08bc",
            
196
      "name": "strategyqa",
237
      "name": "strategyqa",
            
197
      "state": "active",
238
      "state": "active",
            
198
      "vocabulary_id": null
239
      "vocabulary_id": null
            
199
    },
240
    },
            
200
    {
241
    {
            
201
      "display_name": "synthetic",
242
      "display_name": "synthetic",
            
202
      "id": "abe7f9cd-9096-4c14-9cdf-b4b1e273d8d9",
243
      "id": "abe7f9cd-9096-4c14-9cdf-b4b1e273d8d9",
            
203
      "name": "synthetic",
244
      "name": "synthetic",
            
204
      "state": "active",
245
      "state": "active",
            
205
      "vocabulary_id": null
246
      "vocabulary_id": null
            
206
    },
247
    },
            
207
    {
248
    {
            
208
      "display_name": "yes-no questions",
249
      "display_name": "yes-no questions",
            
209
      "id": "5b43d825-ab0b-4700-bbdb-e93c16ca13cd",
250
      "id": "5b43d825-ab0b-4700-bbdb-e93c16ca13cd",
            
210
      "name": "yes-no questions",
251
      "name": "yes-no questions",
            
211
      "state": "active",
252
      "state": "active",
            
212
      "vocabulary_id": null
253
      "vocabulary_id": null
            
213
    }
254
    }
            
214
  ],
255
  ],
            
215
  "title": "StrategyQA",
256
  "title": "StrategyQA",
            
216
  "type": "dataset",
257
  "type": "dataset",
            
217
  "version": ""
258
  "version": ""
            
218
}
259
}