44-EFT.WP.Data.ModelCards v1.0 | 第5章 可选扩展字段

目录 ／ 文档-技术白皮书（V5.05） ／ 44-EFT.WP.Data.ModelCards v1.0

第5章 可选扩展字段

I. 章节目的与范围

II. 分层与触发原则

• 可选（Optional）：当场景需要更高可解释性/可审计性/隐私保障/运维可观测性时启用；启用后须满足本章的类型/正则/依赖锚点与计量校核条款。
• 跨卷依赖：数据契约与导出遵循 Core.DataSpec v1.0；单位/量纲与不确定度遵循 Core.Metrology v1.0；涉及路径/到达时遵循 Core.Equations v1.1；引用一律携带“卷名 vX.Y:锚点”。

III. 可选扩展字段总表

IV. 字段规范与示例

1) explainability

• 语义：可解释性方法、覆盖率与证据工件。
• 结构：
• explainability:
• methods: ["grad", "ig", "lrp"] # Grad/Integrated Gradients/Layer-wise Relevance Prop.
• coverage: {fraction: 0.25, policy: "topk-classes"}
• faithfulness_tests: ["deletion", "insertion"]
• artifacts:
• - {path:"explain/samples/0001.png", sha256:"..."}
• - {path:"explain/report.md", sha256:"..."}

2) compression

• 语义：压缩技术与对精度/延迟/功耗的影响。
• 结构：
• compression:
• pruning: {method:"magnitude", sparsity:0.5}
• quantization: {mode:"int8", scheme:"per-channel", calib:"minmax"}
• distillation: {teacher:"eift.cls.large@v1.2", loss:["ce","mse"]}
• impact:
• accuracy_drop_rel: 0.012
• latency_gain_rel: 0.35
• power_drop_rel: 0.18

3) privacy_preserving

• 语义：差分隐私训练、推理加密与数据最小化策略。
• 结构：
• privacy_preserving:
• training:
• dp: {epsilon: 6.0, delta: 1e-6, accountant:"rdp"}
• inference:
• secure_mode: ["tee","he"] # TEE/同态加密
• pii_scan: true
• data_minimization: true
• notes: "No raw PII retained; hashed identifiers only."

4) audit_trails

• 语义：模型全生命周期审计轨。
• 结构：
• audit_trails:
• lineage:
• code_sha256: "..."
• data_refs: ["eift.obs.demo@v1.0"]
• reports:
• - {path:"eval/summary.csv", sha256:"..."}
• - {path:"robustness/report.md", sha256:"..."}
• dag: {path:"ci/pipeline.dag.json", sha256:"..."}

5) license_constraints

• 语义：许可与使用限制。
• 结构：
• license_constraints:
• license: "Apache-2.0"
• allowed_use: ["academic","benchmark"]
• prohibited_use: ["surveillance","biometric_identification"]
• regional_limits: ["EU-GDPR"]

6) observability

• 语义：线上观测指标与 SLO/SLA。
• 结构：
• observability:
• metrics:
• - {name:"latency_p99_ms", target: 20, window:"5m"}
• - {name:"error_rate", target: 0.005}
• - {name:"drift_kl", target: 0.10}
• alerts:
• - {name:"latency_p99_breach", rule:"latency_p99_ms>20 for 10m", severity:"high"}
• dashboards: ["grafana:board/123"]

7) fallbacks

• 语义：服务降级与兜底策略。
• 结构：
• fallbacks:
• routes:
• - {if:"error_rate>0.01", then:"route_to_baseline"}
• - {if:"latency_p99_ms>50", then:"drop_to_rule_based"}
• baseline_ref: "eift.cls.baseline@v1.0"

8) ablation

• 语义：消融实验与影响归因。
• 结构：
• ablation:
• factors:
• - {name:"augmentation_off", delta_f1_macro:-0.018}
• - {name:"no_bn", delta_f1_macro:-0.037}
• protocol: {seeds:[0,1,2], repeats:3}
• artifacts: [{path:"ablation/table.csv", sha256:"..."}]

V. 机器可读 Schema 片段（规范性）

# I15-5 Optional Extensions (excerpt)

properties:

explainability:

type: object

properties:

methods: {type: array, items: {type: string}}

coverage: {type: object, properties:{fraction:{type:number}, policy:{type:string}}}

faithfulness_tests: {type: array, items:{type:string}}

artifacts: {type: array, items:{type: object, properties:{path:{type:string}, sha256:{type:string}}}}

compression:

type: object

properties:

pruning: {type: object}

quantization: {type: object}

distillation: {type: object}

impact: {type: object, properties:{accuracy_drop_rel:{type:number}, latency_gain_rel:{type:number}, power_drop_rel:{type:number}}}

privacy_preserving:

type: object

properties:

training: {type: object, properties:{dp:{type: object, properties:{epsilon:{type:number}, delta:{type:number}}}}}

inference: {type: object, properties:{secure_mode:{type: array, items:{type:string}}, pii_scan:{type:boolean}}}

data_minimization: {type: boolean}

audit_trails:

type: object

properties:

lineage: {type: object, properties:{code_sha256:{type:string}, data_refs:{type: array, items:{type:string}}}}

reports: {type: array, items:{type: object, properties:{path:{type:string}, sha256:{type:string}}}}

dag: {type: object, properties:{path:{type:string}, sha256:{type:string}}}

license_constraints:

type: object

properties:

license: {type: string}

allowed_use: {type: array, items:{type:string}}

prohibited_use: {type: array, items:{type:string}}

regional_limits: {type: array, items:{type:string}}

observability:

type: object

properties:

metrics: {type: array}

alerts: {type: array}

dashboards: {type: array, items:{type:string}}

fallbacks:

type: object

properties:

routes: {type: array}

baseline_ref: {type: string}

ablation:

type: object

properties:

factors: {type: array}

protocol: {type: object}

artifacts: {type: array}

（Schema 中的引用/工件需在导出清单体现并可校验；量纲由计量章统一校核。）

VI. 与导出清单的耦合（export_manifest 补充）

export_manifest:

artifacts:

- {path:"explain/report.md", sha256:"..."}

- {path:"compression/impact.csv", sha256:"..."}

- {path:"privacy/dp_config.yaml", sha256:"..."}

references:

- "EFT.WP.Core.DataSpec v1.0:EXPORT"

- "EFT.WP.Core.Metrology v1.0:check_dim"

（所有启用的扩展字段，若引入跨卷依赖或新增工件，必须在 references[] 与 artifacts[] 中体现。）

VII. 计量与路径依赖一致性（如适用）

1. 涉及 T_arr 等路径量时，模型卡中仍需登记 delta_form、path="gamma(ell)"、measure="d ell"；两种等价表达并存：
   • T_arr = ( 1 / c_ref ) * ( ∫ n_eff d ell )
   • T_arr = ( ∫ ( n_eff / c_ref ) d ell )
2. 相关数值与单位通过 check_dim 校核；公式/符号/定义禁用中文。

VIII. 本章合规自检

• 启用任一扩展字段时，对应 schema/正则与依赖锚点完整，且通过 Lint。
• see[]/export_manifest.references[] 采用“卷名 vX.Y:锚点”，不使用短码或省略版本。
• 涉及路径量的条目补足 delta_form/path/measure 并通过 check_dim；压缩/隐私条目的影响度量与质量/评测条目一致。
• 数学表达遵循反引号/括号/禁用中文规则。