@researai/deepscientist 1.5.17 → 1.6.0

package/AISB/catalog/aisb.t3.095_fastfeaturecp.yaml

@@ -0,0 +1,153 @@
+id: aisb.t3.095_fastfeaturecp
+name: Fast Feature Conformal Prediction
+version: 0.1.0
+one_line: Accelerate feature-space conformal prediction ~50x via Taylor approximation
+  of nonlinear band transformations.
+task_description: 'This packaged benchmark provides Fast Feature Conformal Prediction
+  (FFCP), a Taylor-expansion approximation of Feature Conformal Prediction (FCP) that
+  significantly reduces computational overhead. FFCP introduces a novel non-conformity
+  score s_ff that avoids expensive nonlinear operations required by FCP to transform
+  confidence bands from feature space to output space. The method is post-hoc, distribution-free,
+  and model-agnostic. It applies to both regression (band length minimization) and
+  classification (prediction set size) tasks. Users calibrate a pretrained base model,
+  then construct conformalized confidence bands with guaranteed coverage at specified
+  confidence levels.
+
+  '
+task_mode: evaluation_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: turnkey
+time_band: 2-6h
+cost_band: medium
+difficulty: medium
+data_access: public
+primary_outputs:
+- band_length
+- coverage_pct
+- conformal_prediction_sets
+launch_profiles:
+- id: ffcp_regression
+  label: FFCP Regression
+  description: 'Run 1D and high-dimensional synthetic regression evaluation. Scripts:
+    FastFeatureCP/main.py. Supports com (1D) and x100-y10-reg (synthetic) datasets.
+
+    '
+- id: ffcp_cqr
+  label: FFCP Quantile Regression (FFCQR)
+  description: 'Run Fast Feature Conformalized Quantile Regression evaluation. Script:
+    FastFeatureCQR/main_FFCQR.py. Combines FFCP concept with quantile regression for
+    asymmetric uncertainty bounds.
+
+    '
+- id: ffcp_classification
+  label: FFCP Classification (FFRAPS)
+  description: 'Run Fast Feature Regularized Adaptive Prediction Sets for ImageNet
+    classification. Script: FastFeatureRAPS/main_FFRaps.py. Requires --data pointing
+    to ImageNet validation set.
+
+    '
+- id: ffcp_segmentation
+  label: FFCP Segmentation
+  description: 'Run Fast Feature Conformal Prediction on Cityscapes for semantic segmentation.
+    Script: FastFeatureCP_seg/main_fcn.py. Requires CITYSCAPES_PATH environment variable.
+
+    '
+dataset_download:
+  primary_method: mixed
+  sources:
+  - type: public_url
+    url: https://deepscientist.cc/AISB/095_fastfeaturecp
+    note: Benchmark archive with bundled synthetic datasets
+  - type: external
+    name: ImageNet Val
+    note: Required for FFRAPS classification variant; user must obtain separately
+  - type: external
+    name: Cityscapes
+    note: Required for segmentation variant; user must obtain separately
+  notes:
+  - Synthetic regression datasets (com, x100-y10-reg) are bundled in the archive
+  - Real-world datasets (ImageNet, Cityscapes) require separate download per their
+    respective licenses
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 50
+    gpu_count: 1
+    gpu_vram_gb: 12
+    gpu_architecture: pascal_or_newer
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 100
+    gpu_count: 1
+    gpu_vram_gb: 24
+    gpu_architecture: ampere_or_newer
+environment:
+  python: '3.8'
+  cuda: '11.6'
+  pytorch: 1.12.1
+  flash_attn: null
+  key_packages:
+  - numpy
+  - pandas
+  - torch
+  - torchvision
+  - scikit-learn
+  - tqdm
+  - matplotlib
+  - pillow
+  notes:
+  - Full dependency set specified in bundled requirements.txt
+  - GPU strongly recommended for segmentation and ImageNet classification variants
+  - CPU-only acceptable for 1D synthetic regression only
+risk_flags:
+- no_runtime_execution_in_packaging_pass
+risk_notes:
+- Static code audit confirmed executable anchors exist for all staged metrics
+- Band_length and coverage_pct backed by FastFeatureCP/main.py
+- Runtime validation recommended before trusting output values
+recommended_when: 'Use this benchmark for post-hoc uncertainty quantification on pretrained
+  models. Ideal when you need (1) faster conformal prediction than standard FCP, (2)
+  guaranteed coverage with shorter prediction bands than Vanilla CP, (3) both regression
+  and classification variants, or (4) model-agnostic calibration without retraining.
+
+  '
+not_recommended_when: 'Do not use this benchmark if you require end-to-end model training
+  from scratch rather than post-hoc conformal calibration. For production deployment,
+  note that Cityscapes and ImageNet variants require separate dataset acquisition.
+
+  '
+paper:
+  title: Accelerating Feature Conformal Prediction via Taylor Approximation
+  venue: NeurIPS 2025
+  year: 2025
+  url: https://arxiv.org/abs/2412.00653
+  github: https://github.com/ElvisWang1111/FastFeatureCP
+  authors:
+  - Zihao Tang
+  - Boyuan Wang
+  - Chuan Wen
+  - Jiaye Teng
+display:
+  palette_seed: silver-teal-conformal
+  art_style: uncertainty-lab
+  accent_priority: high
+image_path: ../image/095_aisb.t3.095_fastfeaturecp.jpg
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.095_fastfeaturecp.zip
+  archive_type: zip
+  local_dir_name: paper-95-FastFeatureCP
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.095_fastfeaturecp.zip
+  sha256: 7c6f8ffd5b2da0cb2f70bf871cedf808a004e8be3b3f3ea68b3dc2a01c6e0471
+  size_bytes: 13655698

package/AISB/catalog/aisb.t3.095_fastfeaturecp.zh.yaml

@@ -0,0 +1,118 @@
+id: aisb.t3.095_fastfeaturecp
+name: 快速特征共形预测
+version: 0.1.0
+one_line: 通过非线性带变换的泰勒近似，将特征空间共形预测加速约 50 倍。
+task_description: 该打包基准测试提供快速特征共形预测 (FFCP)，这是特征共形预测 (FCP) 的泰勒展开近似，可显着减少计算开销。 FFCP 引入了一种新颖的不合格分数 s_ff，避免了 FCP 将置信带从特征空间转换到输出空间所需的昂贵的非线性操作。该方法是事后的、不受分布影响且与模型无关。它适用于回归（带长度最小化）和分类（预测集大小）任务。用户校准预训练的基础模型，然后构建保形置信带，并在指定的置信水平下保证覆盖范围。
+task_mode: evaluation_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: turnkey
+time_band: 2-6h
+cost_band: medium
+difficulty: medium
+data_access: public
+primary_outputs:
+- band_length
+- coverage_pct
+- conformal_prediction_sets
+launch_profiles:
+- id: ffcp_regression
+  label: FFCP回归
+  description: 运行一维和高维综合回归评估。脚本：FastFeatureCP/main.py。支持 com (1D) 和 x100-y10-reg（合成）数据集。
+- id: ffcp_cqr
+  label: FFCP 分位数回归 (FFCQR)
+  description: 运行快速特征保形分位数回归评估。脚本：FastFeatureCQR/main_FFCQR.py。将 FFCP 概念与非对称不确定性边界的分位数回归相结合。
+- id: ffcp_classification
+  label: FFCP 分类 (FFRAPS)
+  description: 运行快速特征正则化自适应预测集以进行 ImageNet 分类。脚本：FastFeatureRAPS/main_FFRaps.py。需要 --data 指向 ImageNet 验证集。
+- id: ffcp_segmentation
+  label: FFCP 细分
+  description: 在城市景观上运行快速特征共形预测以进行语义分割。脚本：FastFeatureCP_seg/main_fcn.py。需要 CITYSCAPES_PATH 环境变量。
+dataset_download:
+  primary_method: mixed
+  sources:
+  - type: public_url
+    url: https://deepscientist.cc/AISB/095_fastfeaturecp
+    note: Benchmark archive with bundled synthetic datasets
+  - type: external
+    name: ImageNet Val
+    note: Required for FFRAPS classification variant; user must obtain separately
+  - type: external
+    name: Cityscapes
+    note: Required for segmentation variant; user must obtain separately
+  notes:
+  - 综合回归数据集（com、x100-y10-reg）捆绑在存档中
+  - 真实世界数据集（ImageNet、Cityscapes）需要根据各自的许可证单独下载
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 50
+    gpu_count: 1
+    gpu_vram_gb: 12
+    gpu_architecture: pascal_or_newer
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 100
+    gpu_count: 1
+    gpu_vram_gb: 24
+    gpu_architecture: ampere_or_newer
+environment:
+  python: '3.8'
+  cuda: '11.6'
+  pytorch: 1.12.1
+  flash_attn: null
+  key_packages:
+  - numpy
+  - pandas
+  - torch
+  - torchvision
+  - scikit-learn
+  - tqdm
+  - matplotlib
+  - pillow
+  notes:
+  - 捆绑的 requirements.txt 中指定的完整依赖项集
+  - 强烈建议使用 GPU 进行分割和 ImageNet 分类变体
+  - 仅 CPU 可接受，仅适用于一维综合回归
+risk_flags:
+- no_runtime_execution_in_packaging_pass
+risk_notes:
+- 静态代码审核确认所有阶段指标都存在可执行锚点
+- Band_length 和 coverage_pct 由 FastFeatureCP/main.py 支持
+- 建议在信任输出值之前进行运行时验证
+recommended_when: 使用此基准对预训练模型进行事后不确定性量化。当您需要 (1) 比标准 FCP 更快的保形预测，(2) 保证覆盖范围且预测带比 Vanilla CP 更短，(3) 回归和分类变体，或 (4) 无需重新训练的模型无关校准时，它是理想的选择。
+not_recommended_when: 如果您需要从头开始进行端到端模型训练而不是事后保形校准，请不要使用此基准。对于生产部署，请注意 Cityscapes 和 ImageNet 变体需要单独的数据集获取。
+paper:
+  title: Accelerating Feature Conformal Prediction via Taylor Approximation
+  venue: NeurIPS 2025
+  year: 2025
+  url: https://arxiv.org/abs/2412.00653
+  github: https://github.com/ElvisWang1111/FastFeatureCP
+  authors:
+  - Zihao Tang
+  - Boyuan Wang
+  - Chuan Wen
+  - Jiaye Teng
+display:
+  palette_seed: silver-teal-conformal
+  art_style: uncertainty-lab
+  accent_priority: high
+image_path: ../image/095_aisb.t3.095_fastfeaturecp.jpg
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.095_fastfeaturecp.zip
+  archive_type: zip
+  local_dir_name: paper-95-FastFeatureCP
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.095_fastfeaturecp.zip
+  sha256: 7c6f8ffd5b2da0cb2f70bf871cedf808a004e8be3b3f3ea68b3dc2a01c6e0471
+  size_bytes: 13655698

package/AISB/catalog/aisb.t3.096_m3svm.yaml

@@ -0,0 +1,189 @@
+id: aisb.t3.096_m3svm
+name: Multi-Class Support Vector Machine with Differential Privacy
+version: 0.1.0
+one_line: Differentially private multi-class SVM using all-in-one single-data-access
+  training with weight and gradient perturbation methods.
+task_description: 'This packaged benchmark covers differentially private multi-class
+  SVMs (PMSVM) that address the privacy-budget scaling problem in traditional OvR
+  and OvO multi-class strategies. Unlike conventional approaches that repeatedly query
+  each data sample when building multiple binary classifiers (consuming the privacy
+  budget proportionally to the number of classes), this method uses all-in-one SVM
+  training to access each data point exactly once. Two perturbation routes are provided:
+  (1) Weight Perturbation (WP), which adds calibrated Gaussian noise to the primal
+  weight matrix of all-in-one SVMs, and (2) Gradient Perturbation (GP), which applies
+  a smoothed hinge-loss approximation with differentially private gradient descent
+  via Opacus. The task preserves rigorous margin guarantees while enabling privacy-preserving
+  multi-class classification with improved utility-privacy trade-offs.
+
+  '
+task_mode: evaluation_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: turnkey
+time_band: 1-2h
+cost_band: low
+difficulty: medium
+data_access: public
+primary_outputs:
+- accuracy
+- accuracy_std
+- private_svm_checkpoint
+launch_profiles:
+- id: gradient_perturbation
+  label: Gradient Perturbation (Opacus)
+  description: 'Run the gradient perturbation route using Opacus for differentially
+    private stochastic gradient descent. Uses smoothed hinge-loss approximation with
+    Adam optimizer. Example: python main-opacus.py --data Cornell --optimizer adam
+    --num_epoch 20 --lr_scale 0.1 --eps 2.0. Supports multiple datasets with dataset-specific
+    hyperparameters. No GPU required but recommended for larger datasets.
+
+    '
+- id: weight_perturbation
+  label: Weight Perturbation (Sklearn)
+  description: 'Run the weight perturbation SVM route using scikit-learn''s LinearSVC
+    with differentially private weight perturbation. Example: python main-sklearn.py
+    --data=Cornell. CPU-only execution is fully supported. Alternative to gradient-based
+    approaches with potentially different utility-privacy trade-offs.
+
+    '
+dataset_download:
+  primary_method: bundled
+  sources:
+  - url: https://deepscientist.cc/AISB/096_m3svm
+    format: zip
+    local_dir: paper-96-M3SVM
+  - source: dataset/
+    format: .mat files
+    description: Preprocessed multi-class datasets (Cornell, Dermatology, Glass, HHAR,
+      ISOLET, ORL, USPS, Vehicle)
+  notes:
+  - Datasets are bundled in .mat format in the dataset/ folder
+  - Dataset-specific hyperparameter configurations are defined in run notebooks and
+    main scripts
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 4
+    ram_gb: 8
+    disk_gb: 10
+    gpu_count: 0
+    gpu_vram_gb: 0
+  recommended:
+    cpu_cores: 8
+    ram_gb: 16
+    disk_gb: 20
+    gpu_count: 1
+    gpu_vram_gb: 8
+  notes:
+  - CPU-only execution is fully viable for both launch profiles
+  - GPU acceleration reduces training time for larger datasets (ISOLET, HHAR) but
+    is not required
+environment:
+  python: '3.8'
+  cuda: null
+  pytorch: 2.0.1
+  flash_attn: null
+  key_packages:
+  - opacus==1.2.0
+  - torch>=2.0.0
+  - scikit-learn>=1.0.0
+  - scipy>=1.7.0
+  - numpy>=1.21.0
+  - pandas>=1.3.0
+  notes:
+  - CPU-only execution is supported and recommended for the minimum configuration
+    route
+  - Opacus is required for the gradient perturbation launch profile only
+  - See bundled requirements.txt for the full dependency set
+  - See bundled README/OPTIMIZATION.md for hyperparameter tuning guidance
+risk_flags:
+- privacy_parameter_tuning
+- hyperparameter_sensitivity
+risk_notes:
+- Privacy budget (epsilon, delta) must be set appropriately for target privacy guarantees
+- Hyperparameter sensitivity: lr_scale, epochs, and C regularization interact with
+    privacy noise
+- Dataset-specific tuning is required; naive defaults may yield poor accuracy
+- Gradient perturbation may require careful clipping threshold selection
+recommended_when: 'Use this benchmark when you need a privacy-aware classical ML task
+  with rigorous differential privacy guarantees. Particularly suitable for tabular
+  multi-class classification with strict data-privacy requirements where traditional
+  deep-learning-heavy benchmarks are overkill. Ideal for evaluating privacy-utility
+  trade-offs in traditional ML settings.
+
+  '
+not_recommended_when: 'Do not use this benchmark if you need large-scale deep learning,
+  multimodal data, sequence modeling, or complex feature extraction pipelines. Not
+  suitable when no privacy requirements exist or when accuracy trade-offs from noise
+  injection are unacceptable for your application.
+
+  '
+paper:
+  title: Multi-Class Support Vector Machine with Differential Privacy
+  venue: NeurIPS
+  year: 2025
+  url: https://openreview.net/forum?id=Jz4PENUsRo
+  authors:
+  - Jinseong Park
+  - Yujin Choi
+  - Jaewook Lee
+  institution:
+  - Korea Institute for Advanced Study
+  - Seoul National University
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.096_m3svm.zip
+  archive_type: zip
+  local_dir_name: paper-96-M3SVM
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.096_m3svm.zip
+  sha256: 320d7f75f10ec480816b2627d4fda2587da3d4455a6119e552fd92b6649e80e9
+  size_bytes: 60022973
+display:
+  palette_seed: ivory-indigo-privacy
+  art_style: classical-ml
+  accent_priority: medium
+image_path: ../image/096_aisb.t3.096_m3svm.jpg
+metric_contract:
+  accuracy:
+    origin_path: run/run_pmsvm_gp-adam.ipynb
+    source_ref: code-backed
+    evaluation_protocol:
+      code_paths:
+      - run/run_pmsvm_gp-adam.ipynb
+      - models/train.py
+      - main-opacus.py
+  accuracy_std:
+    origin_path: run/run_pmsvm_gp-adam.ipynb
+    source_ref: code-backed
+    evaluation_protocol:
+      code_paths:
+      - run/run_pmsvm_gp-adam.ipynb
+      - models/train.py
+  private_svm_checkpoint:
+    origin_path: models/model.py
+    source_ref: code-backed
+    description: Trained model checkpoint with differential privacy applied
+    evaluation_protocol:
+      code_paths:
+      - models/model.py
+      - models/train.py
+capability_tags:
+- research_code_optimization
+- differential_privacy
+- support_vector_machines
+- classification
+- trustworthy_ml
+- privacy_preserving_ml
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+commercial:
+  annual_fee: null

package/AISB/catalog/aisb.t3.096_m3svm.zh.yaml

@@ -0,0 +1,149 @@
+id: aisb.t3.096_m3svm
+name: 具有差分隐私的多类支持向量机
+version: 0.1.0
+one_line: 使用具有权重和梯度扰动方法的一体化单数据访问训练的差分私有多类 SVM。
+task_description: 该打包基准测试涵盖了差分隐私多类 SVM (PMSVM)，可解决传统 OvR 和 OvO 多类策略中的隐私预算扩展问题。与在构建多个二元分类器时重复查询每个数据样本的传统​​方法（消耗与类别数量成比例的隐私预算）不同，该方法使用一体化 SVM 训练来访问每个数据点一次。提供了两种扰动路径：(1) 权重扰动 (WP)，它将校准的高斯噪声添加到一体式 SVM 的原始权重矩阵中；(2) 梯度扰动 (GP)，它通过 Opacus 应用平滑的铰链损失近似和差分私有梯度下降。该任务保留了严格的边际保证，同时通过改进的实用程序与隐私权衡实现保护隐私的多类分类。
+task_mode: evaluation_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: turnkey
+time_band: 1-2h
+cost_band: low
+difficulty: medium
+data_access: public
+primary_outputs:
+- accuracy
+- accuracy_std
+- private_svm_checkpoint
+launch_profiles:
+- id: gradient_perturbation
+  label: 梯度扰动（Opacus）
+  description: 使用 Opacus 运行梯度扰动路线以实现差分私有随机梯度下降。使用带有 Adam 优化器的平滑铰链损失近似。示例： python main-opacus.py --data Cornell --optimizer adam --num_epoch 20 --lr_scale 0.1 --eps 2.0。支持具有特定于数据集的超参数的多个数据集。不需要 GPU，但建议用于较大的数据集。
+- id: weight_perturbation
+  label: 重量扰动 (Sklearn)
+  description: 使用 scikit-learn 的 LinearSVC 和差分私有权重扰动运行权重扰动 SVM 路径。示例： python main-sklearn.py --data=Cornell.完全支持仅 CPU 执行。基于梯度的方法的替代方案，具有潜在不同的效用-隐私权衡。
+dataset_download:
+  primary_method: bundled
+  sources:
+  - url: https://deepscientist.cc/AISB/096_m3svm
+    format: zip
+    local_dir: paper-96-M3SVM
+  - source: dataset/
+    format: .mat files
+    description: Preprocessed multi-class datasets (Cornell, Dermatology, Glass, HHAR, ISOLET, ORL, USPS, Vehicle)
+  notes:
+  - 数据集以 .mat 格式捆绑在 dataset/ 文件夹中
+  - 特定于数据集的超参数配置在运行笔记本和主脚本中定义
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 4
+    ram_gb: 8
+    disk_gb: 10
+    gpu_count: 0
+    gpu_vram_gb: 0
+  recommended:
+    cpu_cores: 8
+    ram_gb: 16
+    disk_gb: 20
+    gpu_count: 1
+    gpu_vram_gb: 8
+  notes:
+  - 仅 CPU 执行对于两种启动配置文件都是完全可行的
+  - GPU 加速可减少较大数据集（ISOLET、HHAR）的训练时间，但不是必需的
+environment:
+  python: '3.8'
+  cuda: null
+  pytorch: 2.0.1
+  flash_attn: null
+  key_packages:
+  - opacus==1.2.0
+  - torch>=2.0.0
+  - scikit-learn>=1.0.0
+  - scipy>=1.7.0
+  - numpy>=1.21.0
+  - pandas>=1.3.0
+  notes:
+  - 支持仅 CPU 执行，并建议采用最低配置路线
+  - 仅梯度扰动发射剖面需要 Opacus
+  - 请参阅捆绑的 requirements.txt 以获取完整的依赖项集
+  - 有关超参数调整指南，请参阅捆绑的 README/OPTIMIZATION.md
+risk_flags:
+- privacy_parameter_tuning
+- hyperparameter_sensitivity
+risk_notes:
+- 必须针对目标隐私保证适当设置隐私预算（epsilon、delta）
+- Hyperparameter sensitivity: lr_scale、纪元和 C 正则化与隐私噪声相互作用
+- 需要针对数据集进行调整；天真的默认值可能会产生较差的准确性
+- 梯度扰动可能需要仔细选择限幅阈值
+recommended_when: 当您需要具有严格差异隐私保证的隐私意识经典 ML 任务时，请使用此基准。特别适合具有严格数据隐私要求的表格多类分类，而传统的深度学习密集型基准则显得有些过分。非常适合评估传统机器学习设置中的隐私与效用权衡。
+not_recommended_when: 如果您需要大规模深度学习、多模态数据、序列建模或复杂的特征提取管道，请勿使用此基准测试。当不存在隐私要求或噪声注入的准确性权衡对于您的应用程序来说是不可接受的时，则不适合。
+paper:
+  title: Multi-Class Support Vector Machine with Differential Privacy
+  venue: NeurIPS
+  year: 2025
+  url: https://openreview.net/forum?id=Jz4PENUsRo
+  authors:
+  - Jinseong Park
+  - Yujin Choi
+  - Jaewook Lee
+  institution:
+  - Korea Institute for Advanced Study
+  - Seoul National University
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.096_m3svm.zip
+  archive_type: zip
+  local_dir_name: paper-96-M3SVM
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.096_m3svm.zip
+  sha256: 320d7f75f10ec480816b2627d4fda2587da3d4455a6119e552fd92b6649e80e9
+  size_bytes: 60022973
+display:
+  palette_seed: ivory-indigo-privacy
+  art_style: classical-ml
+  accent_priority: medium
+image_path: ../image/096_aisb.t3.096_m3svm.jpg
+metric_contract:
+  accuracy:
+    origin_path: run/run_pmsvm_gp-adam.ipynb
+    source_ref: code-backed
+    evaluation_protocol:
+      code_paths:
+      - run/run_pmsvm_gp-adam.ipynb
+      - models/train.py
+      - main-opacus.py
+  accuracy_std:
+    origin_path: run/run_pmsvm_gp-adam.ipynb
+    source_ref: code-backed
+    evaluation_protocol:
+      code_paths:
+      - run/run_pmsvm_gp-adam.ipynb
+      - models/train.py
+  private_svm_checkpoint:
+    origin_path: models/model.py
+    source_ref: code-backed
+    description: 应用差分隐私的训练模型检查点
+    evaluation_protocol:
+      code_paths:
+      - models/model.py
+      - models/train.py
+capability_tags:
+- research_code_optimization
+- differential_privacy
+- support_vector_machines
+- classification
+- trustworthy_ml
+- privacy_preserving_ml
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+commercial:
+  annual_fee: null