Repositories

The Foundry Repository system is an enterprise-grade repository platform for systematically managing AI/ML assets. It provides dedicated repository types for models, datasets, and applications, and supports team collaboration and change history tracking through a Git-like version control system.

Repository Types

Foundry offers three specialized repository types.

Model Repository

A dedicated repository for storing and deploying AI/ML models.

Items you can store:

• Model weight files (PyTorch, TensorFlow, ONNX, etc.)
• Model configuration files (config.json)
• Tokenizer files
• Inference code and example scripts

Dataset Repository

A dedicated repository for managing datasets used for training and evaluation.

Supported Formats:

You can explore data and run SQL queries directly in your web browser via Data Studio.

Space Repository

A repository for hosting interactive applications and demos.

Supported SDKs:

• Gradio (sdk: gradio)
• Docker (sdk: docker) — arbitrary Dockerfile-based apps

A Space determines its build method from the sdk value in the README frontmatter. A container image is built in an isolated build environment and then served.

File Management

Web-Based File Management

• Drag-and-drop file uploads
• Simultaneous upload of multiple files
• Large file handling via S3 Multipart Upload
• File browser with a directory explorer interface

Code Viewer

A syntax-highlighting viewer based on the Monaco editor, supporting over 100 programming languages.

Python SDK

from foundry_hub import FoundryApi

api = FoundryApi(token="your_access_token")
api.upload_file(
    repo_id="org/model-name",
    path_in_repo="model.safetensors",
    path_or_fileobj="./model.safetensors"
)

Version Control (LakeFS)

Foundry provides Git-like version control functionality powered by LakeFS.

Branch Management

• Branch strategy based on the main branch as the default
• Feature branch creation for development work
• Automatic creation of dedicated branches for Pull Requests
• Protection of key branches such as main, master, and develop

Tags and Versioning

• Assign meaningful names to specific commits (e.g., v1.0.0)
• Access specific versions via tags, branches, or commit SHAs
• File comparison between versions

Commit History

• Records commit messages, authors, and timestamps
• Per-file change history tracking
• Diff comparison between commits

Pull Request Collaboration

Pull Requests (PRs) are a core feature for team collaboration, enabling review and discussion of changes.

PR Creation Process

1. Commit your changes (a dedicated PR branch is created automatically)
2. Write a PR title and description
3. Specify the target branch (default: main)
4. Automatic notification sent to reviewers

PR Status:

• Open — Review in progress
• Merged — Merge complete
• Closed — Closed without merging

Code Review

Provides a per-file diff viewer for changes and a comment feature with @mention support.

Merge

Merges using cherry-pick to maintain a linear history. Requires Write permission or higher.

Discussion System

A feature for managing discussions, questions, and issues related to a repository.

• Markdown and @mention support
• Open / Closed status management
• Discussions and Pull Requests share a unified numbering system

Data Studio

A feature for exploring and analyzing dataset repository data directly in your web browser.

Powered by DuckDB WASM, data is processed directly in the browser, enabling large-scale data analysis without any server load.

Key Features:

• Paginated table view (50 rows per page)
• Full-text search and filtering by Split/Subset
• SQL query execution

SELECT * FROM "train" WHERE category = 'positive' LIMIT 100;

• Image/audio media preview

Permission System

Organization Repository Permission Inheritance:

Visibility Settings

• Public — Viewable by all authenticated users
• Private — Accessible only to users with permission

HuggingFace Hub Compatibility

You can use your existing ML workflows as-is simply by setting an environment variable.

import os
os.environ["HF_ENDPOINT"] = "https://your-foundry-instance.example.com"

from huggingface_hub import hf_hub_download
from transformers import AutoModel
from datasets import load_dataset

model_path = hf_hub_download(repo_id="org/model-name", filename="model.safetensors")
model = AutoModel.from_pretrained("org/model-name")
dataset = load_dataset("org/dataset-name")

For migrating from HuggingFace to Foundry, refer to the HuggingFace Import documentation.

Frequently Asked Questions

Q: Is there a file size limit per repository?

There is no limit on individual file sizes; large files can be uploaded via S3 Multipart Upload. Storage capacity is governed by your organization's policy.

Q: How do I invite team members to a private repository?

Create an organization and add your team members as organization members. Repository permissions are granted automatically based on their organization role.

Q: Are there performance issues when analyzing large datasets in Data Studio?

Since data is processed directly on the client using DuckDB WASM, there is no server load. However, for very large datasets, sampling may be applied depending on client memory constraints.

Glossary