mlcommons/croissant

Eclair provides multiple interfaces for dataset discovery and management. Depending on your workflow, you can interact with Eclair using:

• Model Context Protocol (MCP) Tools: For integrating Eclair capabilities into LLM-based agents and environments.
• REST API: For programmatic access to Eclair services over HTTP.
• Python SDK: For direct integration into Python-based machine learning workflows.

The Croissant project involves two distinct licenses:

1. Implementation and Examples: Licensed under Apache 2.
2. Croissant Specification: Licensed under CC BY-ND 4.0.

Usage notes for the Specification (CC BY-ND 4.0):

• You are free to remix and adapt the specification for internal use.
• If you distribute something created based on or adding to the specification, you must reference the specification via a link (do not incorporate the text directly) and include proper attribution.
• You must not change the specification itself.

The Croissant Responsible AI (RAI) specification provides a vocabulary for documenting metadata at two levels:

1. Dataset level: Similar to data cards, these properties describe the overall dataset (e.g., motivation, composition, collection process, and intended uses).
2. Record level: Granular annotations applied to atomic units (e.g., sentences, images, videos) to capture information like specific demographics, concept extraction, or safety/fairness attributes. These record-level insights are often aggregated to create dataset-level 'score cards' or 'nutrition labels'.

Key use cases include documenting the data life cycle, labeling processes, participatory data creation, AI safety/fairness evaluation, traceability for explainability, regulatory compliance, and social/cultural inclusion.

Croissant is a metadata format encoded in JSON-LD designed to standardize the description and semantics of Machine Learning (ML) datasets. It provides a vocabulary for dataset attributes that allows for seamless interchange between ML frameworks like PyTorch, TensorFlow, or JAX.

Key benefits include:

• Discoverability: Enables dataset search engines and LLMs/AI agents to parse metadata for easier discovery.
• Portability and Reproducibility: Provides sufficient information for ML tools to load datasets with minimal code, ensuring identical interpretation across different tools.
• Responsible AI (RAI): Includes machine-actionable mechanisms for data provenance (using W3C PROV-O) and usage conditions (using ODRL) to facilitate responsible data sharing and reuse.

The Croissant RAI specification provides properties to address specific Responsible AI (RAI) use cases. The current version (1.0) focuses on several key areas by extending core Croissant and Schema.org properties with new rai: prefixed properties.

Supported Use Cases

• The data life cycle: Captures how data is collected, its limitations, and how it is maintained.
• Data labeling: Details the platforms, tools, and demographics of annotators.
• Participatory data: Documents the involvement of participants and contributors.
• AI safety and fairness: Addresses biases, limitations, and sensitive information.
• Regulatory compliance: Manages sensitive information and data manipulation protocols.

Note: Use cases for Traceability and Inclusion are identified for future expansion.

Croissant 🥐 is a high-level metadata format for machine learning datasets. It is built on top of schema.org and provides a standardized way to describe datasets so they are easily searchable and machine-readable.

A Croissant file consists of four layers:

1. Metadata: Descriptions of the dataset and responsible ML aspects.
2. Resources: Descriptions of the raw data files or sources.
3. Structure: How raw data is arranged into usable data structures.
4. ML semantics: How the data is typically used in ML contexts.

Eclair employs several testing layers to ensure stability:

• Unit Tests: Test individual components (like EclairTools) in isolation, often using unittest.mock to patch external dependencies.
• Integration Tests: Test interactions between components, such as verifying that the server's /mcp endpoint correctly handles tool calls using httpx.AsyncClient.
• End-to-End (E2E) Tests: Test complete user workflows (e.g., searching for a dataset, downloading it, and retrieving metadata) using the EclairClient.
• Performance Tests: Test the system under load, such as measuring response times for concurrent searches using AsyncEclairClient and asyncio.gather.

The Croissant RAI specification uses several namespaces to extend the core Croissant and Schema.org vocabularies. When constructing your metadata, use these prefixes and IRIs:

GeoCroissant distinguishes between two levels of metadata to support different use cases, such as fairness, safety, and dataset lifecycle management:

1. Dataset-level metadata: Similar to data cards, these provide high-level descriptions of the entire dataset.
2. Record-level metadata: Annotations applied to the atomic units of a dataset (e.g., a single sentence, image, video, or conversation). These granular annotations can be aggregated to provide dataset-level insights, such as coverage of specific topics, concepts, or levels of adversariality for safety testing.