Changes
On December 16, 2024 at 11:05:08 PM UTC, admin:
-
Changed value of field
doi_status
toTrue
in BELLS: A Framework Towards Future Proof Benchmarks for the Evaluation of LLM Safeguards -
Changed value of field
doi_date_published
to2024-12-16
in BELLS: A Framework Towards Future Proof Benchmarks for the Evaluation of LLM Safeguards -
Added resource Original Metadata to BELLS: A Framework Towards Future Proof Benchmarks for the Evaluation of LLM Safeguards
f | 1 | { | f | 1 | { |
2 | "access_rights": "", | 2 | "access_rights": "", | ||
3 | "author": "Diego Dorn", | 3 | "author": "Diego Dorn", | ||
4 | "author_email": "", | 4 | "author_email": "", | ||
5 | "citation": [], | 5 | "citation": [], | ||
6 | "creator_user_id": "17755db4-395a-4b3b-ac09-e8e3484ca700", | 6 | "creator_user_id": "17755db4-395a-4b3b-ac09-e8e3484ca700", | ||
7 | "defined_in": "https://doi.org/10.48550/arXiv.2406.01364", | 7 | "defined_in": "https://doi.org/10.48550/arXiv.2406.01364", | ||
8 | "doi": "10.57702/1yuivokg", | 8 | "doi": "10.57702/1yuivokg", | ||
n | 9 | "doi_date_published": null, | n | 9 | "doi_date_published": "2024-12-16", |
10 | "doi_publisher": "TIB", | 10 | "doi_publisher": "TIB", | ||
n | 11 | "doi_status": false, | n | 11 | "doi_status": true, |
12 | "domain": "https://service.tib.eu/ldmservice", | 12 | "domain": "https://service.tib.eu/ldmservice", | ||
13 | "extra_authors": [ | 13 | "extra_authors": [ | ||
14 | { | 14 | { | ||
15 | "extra_author": "Alexandre Variengien", | 15 | "extra_author": "Alexandre Variengien", | ||
16 | "orcid": "" | 16 | "orcid": "" | ||
17 | }, | 17 | }, | ||
18 | { | 18 | { | ||
19 | "extra_author": "Charbel-Rapha\u00a8el Segerie", | 19 | "extra_author": "Charbel-Rapha\u00a8el Segerie", | ||
20 | "orcid": "" | 20 | "orcid": "" | ||
21 | }, | 21 | }, | ||
22 | { | 22 | { | ||
23 | "extra_author": "Vincent Corruble", | 23 | "extra_author": "Vincent Corruble", | ||
24 | "orcid": "" | 24 | "orcid": "" | ||
25 | } | 25 | } | ||
26 | ], | 26 | ], | ||
27 | "groups": [ | 27 | "groups": [ | ||
28 | { | 28 | { | ||
29 | "description": "", | 29 | "description": "", | ||
30 | "display_name": "Input-Output Safeguards", | 30 | "display_name": "Input-Output Safeguards", | ||
31 | "id": "d7f8087e-aadf-45b5-8c38-12f9fa0262fb", | 31 | "id": "d7f8087e-aadf-45b5-8c38-12f9fa0262fb", | ||
32 | "image_display_url": "", | 32 | "image_display_url": "", | ||
33 | "name": "input-output-safeguards", | 33 | "name": "input-output-safeguards", | ||
34 | "title": "Input-Output Safeguards" | 34 | "title": "Input-Output Safeguards" | ||
35 | }, | 35 | }, | ||
36 | { | 36 | { | ||
37 | "description": "", | 37 | "description": "", | ||
38 | "display_name": "Large Language Models", | 38 | "display_name": "Large Language Models", | ||
39 | "id": "246be871-ce2b-4021-8290-df0f88dec833", | 39 | "id": "246be871-ce2b-4021-8290-df0f88dec833", | ||
40 | "image_display_url": "", | 40 | "image_display_url": "", | ||
41 | "name": "large-language-models", | 41 | "name": "large-language-models", | ||
42 | "title": "Large Language Models" | 42 | "title": "Large Language Models" | ||
43 | } | 43 | } | ||
44 | ], | 44 | ], | ||
45 | "id": "b5d98519-95d8-453e-9d93-c8724a7c9f66", | 45 | "id": "b5d98519-95d8-453e-9d93-c8724a7c9f66", | ||
46 | "isopen": false, | 46 | "isopen": false, | ||
47 | "landing_page": | 47 | "landing_page": | ||
48 | "https://api.semanticscholar.org/CorpusID:269010944", | 48 | "https://api.semanticscholar.org/CorpusID:269010944", | ||
49 | "license_title": null, | 49 | "license_title": null, | ||
50 | "link_orkg": "", | 50 | "link_orkg": "", | ||
51 | "metadata_created": "2024-12-16T23:05:06.957007", | 51 | "metadata_created": "2024-12-16T23:05:06.957007", | ||
n | 52 | "metadata_modified": "2024-12-16T23:05:06.957012", | n | 52 | "metadata_modified": "2024-12-16T23:05:07.494058", |
53 | "name": | 53 | "name": | ||
54 | towards-future-proof-benchmarks-for-the-evaluation-of-llm-safeguards", | 54 | towards-future-proof-benchmarks-for-the-evaluation-of-llm-safeguards", | ||
55 | "notes": "A structured collection of tests for input-output | 55 | "notes": "A structured collection of tests for input-output | ||
56 | safeguards, including established failure tests, emerging failure | 56 | safeguards, including established failure tests, emerging failure | ||
57 | tests, and next-gen architecture tests.", | 57 | tests, and next-gen architecture tests.", | ||
n | 58 | "num_resources": 0, | n | 58 | "num_resources": 1, |
59 | "num_tags": 3, | 59 | "num_tags": 3, | ||
60 | "organization": { | 60 | "organization": { | ||
61 | "approval_status": "approved", | 61 | "approval_status": "approved", | ||
62 | "created": "2024-11-25T12:11:38.292601", | 62 | "created": "2024-11-25T12:11:38.292601", | ||
63 | "description": "", | 63 | "description": "", | ||
64 | "id": "079d46db-32df-4b48-91f3-0a8bc8f69559", | 64 | "id": "079d46db-32df-4b48-91f3-0a8bc8f69559", | ||
65 | "image_url": "", | 65 | "image_url": "", | ||
66 | "is_organization": true, | 66 | "is_organization": true, | ||
67 | "name": "no-organization", | 67 | "name": "no-organization", | ||
68 | "state": "active", | 68 | "state": "active", | ||
69 | "title": "No Organization", | 69 | "title": "No Organization", | ||
70 | "type": "organization" | 70 | "type": "organization" | ||
71 | }, | 71 | }, | ||
72 | "owner_org": "079d46db-32df-4b48-91f3-0a8bc8f69559", | 72 | "owner_org": "079d46db-32df-4b48-91f3-0a8bc8f69559", | ||
73 | "private": false, | 73 | "private": false, | ||
74 | "relationships_as_object": [], | 74 | "relationships_as_object": [], | ||
75 | "relationships_as_subject": [], | 75 | "relationships_as_subject": [], | ||
t | 76 | "resources": [], | t | 76 | "resources": [ |
77 | { | ||||
78 | "cache_last_updated": null, | ||||
79 | "cache_url": null, | ||||
80 | "created": "2024-12-16T23:37:09", | ||||
81 | "data": [ | ||||
82 | "dcterms:title", | ||||
83 | "dcterms:accessRights", | ||||
84 | "dcterms:creator", | ||||
85 | "dcterms:description", | ||||
86 | "dcterms:issued", | ||||
87 | "dcterms:language", | ||||
88 | "dcterms:identifier", | ||||
89 | "dcat:theme", | ||||
90 | "dcterms:type", | ||||
91 | "dcat:keyword", | ||||
92 | "dcat:landingPage", | ||||
93 | "dcterms:hasVersion", | ||||
94 | "dcterms:format", | ||||
95 | "mls:task", | ||||
96 | "datacite:isDescribedBy" | ||||
97 | ], | ||||
98 | "description": "The json representation of the dataset with its | ||||
99 | distributions based on DCAT.", | ||||
100 | "format": "JSON", | ||||
101 | "hash": "", | ||||
102 | "id": "38826a2c-3749-4f86-aa93-6a23b0636ce2", | ||||
103 | "last_modified": "2024-12-16T23:05:07.486286", | ||||
104 | "metadata_modified": "2024-12-16T23:05:07.497217", | ||||
105 | "mimetype": "application/json", | ||||
106 | "mimetype_inner": null, | ||||
107 | "name": "Original Metadata", | ||||
108 | "package_id": "b5d98519-95d8-453e-9d93-c8724a7c9f66", | ||||
109 | "position": 0, | ||||
110 | "resource_type": null, | ||||
111 | "size": 921, | ||||
112 | "state": "active", | ||||
113 | "url": | ||||
114 | resource/38826a2c-3749-4f86-aa93-6a23b0636ce2/download/metadata.json", | ||||
115 | "url_type": "upload" | ||||
116 | } | ||||
117 | ], | ||||
77 | "services_used_list": "", | 118 | "services_used_list": "", | ||
78 | "state": "active", | 119 | "state": "active", | ||
79 | "tags": [ | 120 | "tags": [ | ||
80 | { | 121 | { | ||
81 | "display_name": "Large Language Models", | 122 | "display_name": "Large Language Models", | ||
82 | "id": "6e366b4a-402d-4c0e-b818-6fec429fc72f", | 123 | "id": "6e366b4a-402d-4c0e-b818-6fec429fc72f", | ||
83 | "name": "Large Language Models", | 124 | "name": "Large Language Models", | ||
84 | "state": "active", | 125 | "state": "active", | ||
85 | "vocabulary_id": null | 126 | "vocabulary_id": null | ||
86 | }, | 127 | }, | ||
87 | { | 128 | { | ||
88 | "display_name": "benchmarking", | 129 | "display_name": "benchmarking", | ||
89 | "id": "d9f3d223-10a3-41fb-9231-bed79d8943c0", | 130 | "id": "d9f3d223-10a3-41fb-9231-bed79d8943c0", | ||
90 | "name": "benchmarking", | 131 | "name": "benchmarking", | ||
91 | "state": "active", | 132 | "state": "active", | ||
92 | "vocabulary_id": null | 133 | "vocabulary_id": null | ||
93 | }, | 134 | }, | ||
94 | { | 135 | { | ||
95 | "display_name": "input-output safeguards", | 136 | "display_name": "input-output safeguards", | ||
96 | "id": "ac31fd55-69c7-485a-9781-79042fc413c4", | 137 | "id": "ac31fd55-69c7-485a-9781-79042fc413c4", | ||
97 | "name": "input-output safeguards", | 138 | "name": "input-output safeguards", | ||
98 | "state": "active", | 139 | "state": "active", | ||
99 | "vocabulary_id": null | 140 | "vocabulary_id": null | ||
100 | } | 141 | } | ||
101 | ], | 142 | ], | ||
102 | "title": "BELLS: A Framework Towards Future Proof Benchmarks for the | 143 | "title": "BELLS: A Framework Towards Future Proof Benchmarks for the | ||
103 | Evaluation of LLM Safeguards", | 144 | Evaluation of LLM Safeguards", | ||
104 | "type": "dataset", | 145 | "type": "dataset", | ||
105 | "version": "" | 146 | "version": "" | ||
106 | } | 147 | } |