File size: 20,871 Bytes
a8af1a7
 
 
7c15f75
 
 
 
 
 
a8af1a7
 
7c15f75
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a8af1a7
 
7c15f75
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a8af1a7
 
6301ef2
7c15f75
 
 
 
 
 
 
 
 
a8af1a7
 
7c15f75
 
 
 
 
 
 
 
 
a8af1a7
 
7c15f75
 
 
 
 
 
 
 
 
 
a8af1a7
 
7c15f75
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a8af1a7
 
7c15f75
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a8af1a7
 
 
7c15f75
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
{
  "metadata": {
    "Name": "Model C",
    "Provider": "BigCode",
    "URL": "https://huggingface.co/bigcode/starcoder2-15b",
    "Type": "Large Language Model",
    "Modalities": [
      "Text-to-Text"
    ]
  },
  "scores": {
    "1. Bias, Stereotypes, and Representational Harms Evaluation": {
      "1.1 Bias Detection Overview": {
        "status": "Yes",
        "sources": [
          {
            "type": "🌐",
            "detail": "https://arxiv.org/abs/2402.19173",
            "name": "BOLD - Bias in Open-ended Language Generation Dataset"
          },
          {
            "type": "🌐",
            "detail": "https://arxiv.org/abs/2402.19173",
            "name": "WinoBias"
          }
        ],
        "questions": {
          "Evaluations at various stages (data collection, preprocessing, AI system architecture, training, deployment)": false,
          "Have intrinsic properties of the AI system been evaluated for bias (e.g., embedding analysis)": false,
          "Have extrinsic bias evaluations been run (e.g., downstream task performance)": true,
          "Have evaluations been run across all applicable modalities": true,
          "Have bias evaluations been run that take the form of automatic quantitative evaluation": true,
          "Have bias evaluations been run with human participants?": false
        }
      },
      "1.2 Protected Classes and Intersectional Measures": {
        "status": "No",
        "sources": [],
        "questions": {
          "Do evaluations cover all applicable legal protected categories for in-scope uses of the system?": false,
          "Do evaluations cover additional subgroups that are likely to be harmed based on other personal characteristics": false,
          "Evaluation of how different aspects of identity interact and compound in AI system behavior": false,
          "Evaluation of AI system biases for legal protected categories and additional relevant subgroups": false
        }
      },
      "1.3 Measurement of Stereotypes and Harmful Associations": {
        "status": "Yes",
        "sources": [
          {
            "type": "🌐",
            "detail": "https://arxiv.org/abs/2402.19173",
            "name": "HONEST - Hurtful Sentence Completion in English Language Models"
          },
          {
            "type": "🌐",
            "detail": "https://arxiv.org/abs/2402.19173",
            "name": "RealToxicityPrompts"
          }
        ],
        "questions": {
          "Measurement of known stereotypes in AI system outputs": true,
          "Measurement of other negative associations and assumptions regarding specific groups": true,
          "Measurement of stereotypes and negative associations across in-scope contexts": false
        }
      },
      "1.4 Bias Evaluation Transparency and Documentation": {
        "status": "Yes",
        "sources": [
          {
            "type": "🌐",
            "detail": "https://arxiv.org/abs/2402.19173",
            "name": "Evaluation Documentation"
          }
        ],
        "questions": {
          "Sufficient documentation of evaluation methods (including code and datasets) to replicate findings": true,
          "Sufficient documentation of evaluation results (including intermediary statistics) to support comparison to other AI systems": true,
          "Documentation of bias mitigation measures, including their secondary impacts": false,
          "Documentation of bias monitoring approaches post-release/deployment if applicable": false
        }
      }
    },
    "2. Cultural Values and Sensitive Content Evaluation": {
      "2.1 Cultural Variation Overview": {
        "status": "N/A",
        "sources": [],
        "questions": {
          "Evaluations at various stages (data collection, preprocessing, AI system architecture, training, deployment)": false,
          "Have intrinsic properties of the AI system been evaluated for cultural variation(e.g., embedding analysis)": false,
          "Have extrinsic cultural variation evaluations been run (e.g., downstream task performance)": false,
          "Have evaluations been run across all applicable modalities": false,
          "Have cultural variation evaluations been run that take the form of automatic quantitative evaluation": false,
          "Have cultural variation evaluations been run with human participants?": false
        }
      },
      "2.2 Cultural Diversity and Representation": {
        "status": "N/A",
        "sources": [],
        "questions": {
          "Use of evaluation methods developed in the cultural contexts in scope": false,
          "Respect of indigenous sovereignty, protected rights, and cultural norms in AI system-generated content": false,
          "Evaluation of cultural variation across geographic dimensions": false,
          "Evaluation of cultural variation representing communities' perspectives within geographical contexts": false,
          "Analysis of how cultural context affects AI system performance": false
        }
      },
      "2.3 Generated Sensitive Content across Cultural Contexts": {
        "status": "Yes",
        "sources": [
          {
            "type": "🌐",
            "detail": "https://arxiv.org/abs/2402.19173",
            "name": "HONEST - Hurtful Sentence Completion in English Language Models"
          },
          {
            "type": "🌐",
            "detail": "https://arxiv.org/abs/2402.19173",
            "name": "RealToxicityPrompts"
          }
        ],
        "questions": {
          "Has the AI system been evaluated for its likelihood of facilitating generation of threatening or violent content": true,
          "Has the AI system been evaluated for its likelihood of facilitating generation of targeted harassment or discrimination": false,
          "Has the AI system been evaluated for its likelihood of facilitating generation of hate speech": false,
          "Has the AI system been evaluated for its likelihood of exposing its direct users to content embedding values and assumptions not reflective of their cultural context": false,
          "Has the AI system been evaluated for its likelihood of exposing its direct users to inappropriate content for their use context": true,
          "Has the AI system been evaluated for its likelihood of exposing its direct users to content with negative psychological impacts": false,
          "Has the evaluation of the AI system's behaviors explicitly considered cultural variation in their definition": false
        }
      },
      "2.4 Cultural Variation Transparency and Documentation": {
        "status": "N/A",
        "sources": [],
        "questions": {
          "Documentation of cultural contexts considered during development": false,
          "Documentation of the range of cultural contexts covered by evaluations": false,
          "Sufficient documentation of evaluation method to understand the scope of the findings": false,
          "Construct validity, documentation of strengths, weaknesses, and assumptions": false,
          "Domain shift between evaluation development and AI system development settings": false,
          "Sufficient documentation of evaluation methods to replicate findings": false,
          "Sufficient documentation of evaluation results to support comparison": false,
          "Document of psychological impact on evaluators reviewing harmful content": false,
          "Documentation of measures to protect evaluator well-being": false
        }
      }
    },
  "3. Disparate Performance Evaluation": {
    "3.1 Disparate Performance Overview": {
      "status": "N/A",
      "sources": [],
      "questions": {
        "Have development choices and intrinsic properties of the AI system been evaluated for their contribution to disparate performance?": false,
        "Have extrinsic disparate performance evaluations been run": false,
        "Have evaluations been run across all applicable modalities": false,
        "Have disparate performance evaluations been run that take the form of automatic quantitative evaluation": false,
        "Have disparate performance evaluations been run with human participants": false
      }
    },
    "3.2 Identifying Target Groups for Disparate Performance Evaluation": {
      "status": "N/A",
      "sources": [],
      "questions": {
        "Identification of mandated target group based on legal nondiscrimination frameworks": false,
        "Identification of further target groups that are likely to be harmed by disparate performance": false,
        "Assessment of systemic barriers in dataset collection methods for different groups": false,
        "Consideration of historical disparities in the task in which the AI system is deployed": false,
        "Identification of both implicit and explicit markers for the target groups": false
      }
    },
    "3.3 Subgroup Performance Analysis": {
      "status": "N/A",
      "sources": [],
      "questions": {
        "Non-aggregated evaluation results across subpopulations, including feature importance and consistency analysis": false,
        "Metrics to measure performance in decision-making tasks": false,
        "Metrics to measure disparate performance in other tasks including generative tasks": false,
        "Worst-case subgroup performance analysis, including performance on rare or underrepresented cases": false,
        "Intersectional analysis examining performance across combinations of subgroup": false,
        "Do evaluations of disparate performance account for implicit social group markers": false
      }
    },
    "3.4 Disparate Performance Evaluation Transparency and Documentation": {
      "status": "N/A",
      "sources": [],
      "questions": {
        "Sufficient documentation of evaluation method to understand the scope of the findings": false,
        "Documentation of strengths, weaknesses, and assumptions about the context": false,
        "Documentation of domain shift between evaluation and deployment settings": false,
        "Sufficient documentation of evaluation methods to replicate findings": false,
        "Sufficient documentation of evaluation results to support comparison": false,
        "Documentation of disparate performance mitigation measures": false,
        "Documentation of disparate performance monitoring approaches": false
      }
    }
  },
  "4. Environmental Costs and Carbon Emissions Evaluation": {
    "4.1 Environmental Costs Overview": {
      "status": "Yes",
      "sources": [
        {
          "type": "🌐",
          "detail": "https://mlco2.github.io/impact/#compute",
          "name": "Machine Learning Emissions Calculator"
        }
      ],
      "questions": {
        "Evaluations of different processes within development and deployment": false,
        "Have evaluations been run across all applicable modalities?": true,
        "Have evaluations been run on standardized benchmarks or metrics?": true,
        "Have evaluations taken into account community feedback from regions affected by data center power consumption?": false,
        "Do evaluations consider the full supply chain including environmental impact of hardware components and data centers used?": false
      }
    },
    "4.2 Energy Cost and Environmental Impact of Development": {
      "status": "Yes",
      "sources": [
        {
          "type": "🌐",
          "detail": "https://mlco2.github.io/impact/#compute",
          "name": "Machine Learning Emissions Calculator"
        }
      ],
      "questions": {
        "Accounting of FLOPS across development stages": true,
        "Evaluation of energy consumption using standardized tracking tools": true,
        "Evaluation of carbon impact accounting for regional energy sources": true,
        "Evaluation of hardware lifecycle environmental impact": false
      }
    },
    "4.3 Energy Cost and Environmental Impact of Deployment": {
      "status": "N/A",
      "sources": [],
      "questions": {
        "Evaluation of inference FLOPS for the system": false,
        "Evaluation of inference energy consumption on most common deployment setting": false,
        "Evaluation of inference energy consumption on multiple deployment settings": false,
        "Evaluation of task-specific energy consumption variations": false,
        "Evaluation of carbon impact for deployment infrastructure": false,
        "Evaluation of hardware lifecycle environmental impact for deployment": false
      }
    },
    "4.4 Environmental Costs Transparency and Documentation": {
      "status": "Yes",
      "sources": [
        {
          "type": "🌐",
          "detail": "https://mlco2.github.io/impact/#compute",
          "name": "Machine Learning Emissions Calculator"
        }
      ],
      "questions": {
        "Documentation about equipment and infrastructure specifications": true,
        "Sufficient documentation of evaluation methods including components covered": false,
        "Sufficient documentation of evaluation methods to replicate findings": true,
        "Sufficient documentation of evaluation results for comparison": true
      }
    }
  },
  "5. Privacy and Data Protection Evaluation": {
    "5.1 Privacy and Data Protection Overview": {
      "status": "Yes",
      "sources": [
        {
          "type": "🏒",
          "detail": "PII detection and redaction using an NER model"
        },
        {
          "type": "🌐",
          "detail": "https://huggingface.co/spaces/bigcode/in-the-stack",
          "name": "Opt-out tool for users"
        },
        {
          "type": "🌐",
          "detail": "https://arxiv.org/abs/2402.19173",
          "name": "Asleep at the Keyboard Security Benchmark"
        }
      ],
      "questions": {
        "Evaluations at various stages (data collection, preprocessing, AI system architecture, training, deployment)": true,
        "Have intrinsic properties of the AI system been evaluated for privacy vulnerabilities": false,
        "Have extrinsic privacy evaluations been run": true,
        "Have evaluations been run across all applicable modalities": true,
        "Have privacy evaluations been run that take the form of automatic quantitative evaluation": true,
        "Have privacy evaluations been run with human participants?": false
      }
    },
    "5.2 Privacy, Likeness, and Publicity Harms": {
      "status": "N/A",
      "sources": [],
      "questions": {
        "Has the AI system been evaluated for its likelihood of revealing personal information from its training data?": false,
        "Has the AI system been evaluated for its likelihood of facilitating generation of content impersonating an individual?": false,
        "Has the AI system been evaluated for its likelihood of providing made up or confabulated personal information about individuals?": false
      }
    },
    "5.3 Intellectual Property and Information Security": {
      "status": "Yes",
      "sources": [
        {
          "type": "🏒",
          "detail": "Membership test to find if generated code was copied from the training corpus"
        },
        {
          "type": "🏒",
          "detail": "Code attribution tool to find the original author and license of the generated code"
        },
        {
          "type": "🌐",
          "detail": "https://arxiv.org/abs/2402.19173",
          "name": "Asleep at the Keyboard Security Benchmark"
        }
      ],
      "questions": {
        "Has the AI system been evaluated for its likelihood of reproducing other categories of information from its training data": true,
        "Has the system been evaluated for other information security risks for in-scope uses": false
      }
    },
    "5.4 Privacy Evaluation Transparency and Documentation": {
      "status": "Yes",
      "sources": [
        {
          "type": "🏒",
          "detail": "Documentation of training data information risk categories and consent status"
        }
      ],
      "questions": {
        "Documentation of the categories of training data that present information risk": true,
        "Documentation of evaluation methods to replicate findings": true,
        "Documentation of evaluation results to support comparison": true,
        "Documentation of evaluation limitations": false,
        "Documentation of deployment considerations": false
      }
    }
  },
  "6. Financial Costs Evaluation": {
    "6.1 Financial Costs Overview": {
      "status": "N/A",
      "sources": [],
      "questions": {
        "Evaluation of costs at various stages": false,
        "Have costs been evaluated for different system components": false,
        "Have cost evaluations been run across all applicable modalities": false,
        "Have cost evaluations included both direct and indirect expenses": false,
        "Have cost projections been validated against actual expenses": false
      }
    },
    "6.2 Development and Training Costs": {
      "status": "N/A",
      "sources": [],
      "questions": {
        "Assessment of research and development labor costs": false,
        "Evaluation of data collection and preprocessing costs": false,
        "Assessment of training infrastructure costs": false,
        "Assessment of costs associated with different training approaches": false,
        "Evaluation of model architecture and size impact on costs": false
      }
    },
    "6.3 Deployment and Operation Costs": {
      "status": "N/A",
      "sources": [],
      "questions": {
        "Assessment of inference and serving costs": false,
        "Evaluation of storage and hosting expenses": false,
        "Assessment of scaling costs based on usage patterns": false,
        "Evaluation of costs specific to different deployment contexts": false,
        "Assessment of costs for model updates or fine-tuning by end users": false
      }
    },
    "6.4 Financial Cost Documentation and Transparency": {
      "status": "N/A",
      "sources": [],
      "questions": {
        "Sufficient documentation of cost evaluation methodology and assumptions": false,
        "Sufficient documentation of cost breakdowns and metrics": false,
        "Documentation of cost variations across different usage scenarios": false,
        "Documentation of long-term cost projections and risk factors": false
      }
    }
  },
  "7. Data and Content Moderation Labor Evaluation": {
    "7.1 Labor Evaluation Overview": {
      "status": "Yes",
      "sources": [
        {
          "type": "🏒",
          "detail": "PII annotations by human annotators with fair wage"
        }
      ],
      "questions": {
        "Evaluation of labor practices at various stages": true,
        "Have labor conditions been evaluated for different worker categories": true,
        "Have labor evaluations been run across all applicable task types": false,
        "Have labor practices been evaluated against established industry standards": true,
        "Have labor evaluations included both direct employees and contracted workers": false,
        "Have evaluations considered different regional and jurisdictional contexts": true
      }
    },
    "7.2 Working Conditions and Compensation": {
      "status": "Yes",
      "sources": [
        {
          "type": "🏒",
          "detail": "PII annotations by human annotators with fair wage"
        }
      ],
      "questions": {
        "Assessment of compensation relative to local living wages and industry standards": true,
        "Assessment of job security and employment classification": false,
        "Evaluation of workplace safety, worker protections and rights": false,
        "Assessment of worker autonomy and task assignment practices": false,
        "Evaluation of power dynamics and worker feedback mechanisms": false
      }
    },
    "7.3 Worker Wellbeing and Support": {
      "status": "N/A",
      "sources": [],
      "questions": {
        "Assessment of psychological support systems, trauma resources, and other long-term mental health monitoring": false,
        "Evaluation of training and preparation for difficult content": false,
        "Evaluation of cultural and linguistic support for diverse workforces": false
      }
    },
    "7.4 Labor Practice Documentation and Transparency": {
      "status": "Yes",
      "sources": [
        {
          "type": "🏒",
          "detail": "PII annotations by human annotators with fair wage"
        }
      ],
      "questions": {
        "Documentation of labor evaluation methodology and frameworks used": true,
        "Documentation of worker demographics and task distribution": false,
        "Documentation of support systems, worker protections": false,
        "Documentation of incident reporting and resolution procedures": false
      }
    }
  }
}
}