Selecting the right modeling tool（选择合适的建模工具）¶
Palantir's modeling suite of products enables users to develop, manage, and operationalize models. This page compares different products to help you choose the right tool for your needs.
For guided assistance with modeling tasks, AI FDE offers a dedicated Machine learning mode that helps you train, evaluate, deploy, and tune models. The agent can guide you through the full workflow, from feature engineering through model training and deployment, using either Model Studio or pro-code repositories. To get started, select the machine learning mode from the AI FDE mode selector or describe your modeling task and let the agent select the mode automatically.
Feature engineering¶
| Product | Details |
|---|---|
| Pipeline Builder | Large scale point-and-click data transformation |
| Code Workspaces | Interactive, pro-code data analysis and transformation in familiar environments such as JupyterLab® |
| Python Transforms | PySpark data pipeline development in Foundry's web-based IDE, Code Repositories |
No-code model training¶
No-code model training tools are available in Model Studio, providing a simple point-and-click interface for creating production-grade machine learning models.
Pro-code model training¶
Available libraries¶
The palantir_models library provides flexible tooling to publish and consume models within the Palantir platform, using the concept of model adapters. The foundry_ml library, its predecessor, has been formally deprecated as of October 2025.
Code authoring environments¶
| Product | Library support | Details |
|---|---|---|
| Code Workspaces | palantir_models |
Interactive model development in Jupyter® notebooks |
| Code Repositories | palantir_models |
Powerful web-based IDE with native CI/CD features and support for modeling workflows; less interactive than notebooks |
Training metrics tracking¶
| Product | Details |
|---|---|
| Experiments | Framework for logging metrics and hyperparameters during a model training job |
Batch inference¶
Models can be used for running large scale batch inference pipelines across datasets.
| Product | Details | Caveats |
|---|---|---|
| Pipeline Builder | No-code model inference directly on the pipeline canvas using the trained model node. Models run as isolated sidecars alongside Spark executors and automatically use the latest model version. | Only supports models with a single tabular input and output. Streaming and Lightweight execution modes are not yet supported. |
| Python transforms | Batch inference can be run directly in Python transforms. Supports pinning a specific model version. | Using the @lightweight decorator and model sidecars is recommended. |
| Modeling objective batch deployments | Modeling Objectives offers broader model management features such as model release management and evaluation. | Does not support multi-output and external models, models as sidecars, or deployment via Marketplace as detailed here. |
| Jupyter® Notebook | Users can create scheduled training and/or inference jobs directly from Code Workspaces. | Only supports running inference models created from the same notebook; use Python Transforms to orchestrate models created elsewhere. |
Model deployment¶
Models can be deployed in Foundry behind a REST API; deploying a model operationalizes the model for use both inside and outside of Foundry.
| Product | Details |
|---|---|
| Model direct deployments | Auto-upgrading model deployments; best for quick iteration and deployment. |
| Modeling objective live deployments | Production-grade modeling project management; modeling objectives provide tooling for model release management and evaluation. Does not support deployment via Marketplace as detailed here. |
Functions integration¶
Publishing models as functions makes it easy to use models for live inference in downstream Foundry applications, including Workshop, Slate, actions, and more.
| Product | Best for |
|---|---|
| Direct function publication | No-code function creation on models with live deployments, allowing integration with the Ontology. The same functionality is available in the Model and Modeling Objectives applications. |
| Importing model functions into Functions repositories | Import model functions into TypeScript v1, v2 or Python functions to further process predictions (for example, make ontology edits) with support for Model API type checking. |
中文翻译¶
选择合适的建模工具¶
Palantir的建模产品套件使用户能够开发、管理和运维模型。本页面通过比较不同产品，帮助您根据需求选择最合适的工具。
如需建模任务的引导式协助，AI FDE 提供了专门的机器学习模式，可帮助您训练、评估、部署和调优模型。该智能体能够引导您完成从特征工程到模型训练和部署的完整工作流，既可使用 Model Studio，也可使用专业代码仓库。要开始使用，请从AI FDE模式选择器中选择机器学习模式，或描述您的建模任务，让智能体自动选择模式。
特征工程¶
| 产品 | 详情 |
|---|---|
| Pipeline Builder | 大规模点击式数据转换 |
| Code Workspaces | 在JupyterLab®等熟悉环境中进行交互式专业代码数据分析与转换 |
| Python Transforms | 在Foundry基于Web的IDE Code Repositories 中进行PySpark数据管道开发 |
无代码模型训练¶
无代码模型训练工具可在 Model Studio 中使用，提供简单的点击式界面，用于创建生产级机器学习模型。
专业代码模型训练¶
可用库¶
palantir_models 库提供了灵活的工具，用于在Palantir平台内发布和使用模型，其核心概念是模型适配器(Model Adapters)。其前身 foundry_ml 库已于 2025年10月 正式弃用。
代码编写环境¶
| 产品 | 库支持 | 详情 |
|---|---|---|
| Code Workspaces | palantir_models |
在Jupyter®笔记本中进行交互式模型开发 |
| Code Repositories | palantir_models |
功能强大的基于Web的IDE，具备原生CI/CD功能并支持建模工作流；交互性不如笔记本 |
训练指标追踪¶
| 产品 | 详情 |
|---|---|
| Experiments | 用于在模型训练任务期间记录指标和超参数的框架 |
批量推理¶
模型可用于跨数据集运行大规模批量推理管道。
| 产品 | 详情 | 注意事项 |
|---|---|---|
| Pipeline Builder | 使用训练模型节点(Trained Model Node)在管道画布上直接进行无代码模型推理。模型作为独立边车(Sidecar)与Spark执行器一起运行，并自动使用最新模型版本。 | 仅支持具有单一表格输入和输出的模型。尚不支持流式执行模式和轻量级执行模式。 |
| Python transforms | 可直接在Python转换中运行批量推理。支持固定特定模型版本。 | 建议使用 @lightweight 装饰器 和模型边车(Model Sidecars)。 |
| 建模目标(Modeling Objective) 批量部署 | 建模目标提供更广泛的模型管理功能，如模型发布管理和评估。 | 不支持多输出和外部模型、模型作为边车，或通过Marketplace进行部署，详见此处。 |
| Jupyter® Notebook | 用户可以直接从 Code Workspaces 创建定时训练和/或推理任务。 | 仅支持运行从同一笔记本创建的推理模型；如需编排在其他地方创建的模型，请使用Python Transforms。 |
模型部署¶
模型可以在Foundry中通过REST API进行部署；部署模型可使模型在Foundry内外均可投入生产使用。
| 产品 | 详情 |
|---|---|
| 模型直接部署(Model Direct Deployments) | 自动升级的模型部署；最适合快速迭代和部署。 |
| 建模目标(Modeling Objective) 在线部署(Live Deployments) | 生产级建模项目管理；建模目标提供模型发布管理和评估工具。不支持通过Marketplace进行部署，详见此处。 |
函数集成¶
将模型发布为函数，可以方便地在下游Foundry应用程序中使用模型进行实时推理，包括 Workshop、Slate、操作(Actions) 以及更多。
| 产品 | 最佳适用场景 |
|---|---|
| 直接函数发布(Direct Function Publication) | 对已部署的模型进行无代码函数创建，允许与本体论(Ontology)集成。模型和建模目标应用程序中也提供相同的功能。 |
| 将模型函数导入函数仓库 | 将模型函数导入TypeScript v1、v2或Python函数，以进一步处理预测结果（例如，进行本体编辑），并支持模型API类型检查。 |