model_data/model_a_data.json

{
  "metadata": {
    "Name": "StarCoder2",
    "Provider": "BigCode",
    "URL": "https://huggingface.co/bigcode/starcoder2-15b",
    "Type": "Large Language Model",
    "Modalities": ["Text-to-Text"]
  },
  "scores": {
    "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
        }
      }
    },
    "Cultural Values and Sensitive Content Evaluation": {
      "2.1 Cultural Variation Overview": {
        "status": "N/A",
        "sources": [],
        "questions": {}
      },
      "2.2 Cultural Diversity and Representation": {
        "status": "N/A",
        "sources": [],
        "questions": {}
      },
      "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": false,
          "Has the AI system been evaluated for its likelihood of exposing its direct users to categories of content that might be inappropriate": true,
          "Has the AI system been evaluated for its likelihood of exposing its direct users to content that might have additional negative psychological impacts": false,
          "Has the evaluation of the AI system's behaviors explicitly considered cultural variation": false
        }
      },
      "2.4 Cultural Variation Transparency and Documentation": {
        "status": "N/A",
        "sources": [],
        "questions": {}
      }
    },
    "Disparate Performance": {
      "3.1 Disparate Performance Overview": {
        "status": "N/A",
        "sources": [],
        "questions": {}
      },
      "3.2 Identifying Target Groups for Disparate Performance Evaluation": {
        "status": "N/A",
        "sources": [],
        "questions": {}
      },
      "3.3 Subgroup Performance Analysis": {
        "status": "N/A",
        "sources": [],
        "questions": {}
      },
      "3.4 Disparate Performance Evaluation Transparency and Documentation": {
        "status": "N/A",
        "sources": [],
        "questions": {}
      }
    },
    "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": {}
      },
      "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
        }
      }
    },
    "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": {}
      },
      "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
        }
      }
    },
    "Financial Costs Evaluation": {
      "6.1 Financial Costs Overview": {
        "status": "N/A",
        "sources": [],
        "questions": {}
      },
      "6.2 Development and Training Costs": {
        "status": "N/A",
        "sources": [],
        "questions": {}
      },
      "6.3 Deployment and Operation Costs": {
        "status": "N/A",
        "sources": [],
        "questions": {}
      },
      "6.4 Financial Cost Documentation and Transparency": {
        "status": "N/A",
        "sources": [],
        "questions": {}
      }
    },
    "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": {}
      },
      "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
        }
      }
    }
  }
}