Foundation Model Transparency Report

A high-level description of the datasets used for pre-training, including modalities (text, image, code), data sources (web crawl, licensed repositories, synthetic), and temporal cut-off dates. It must disclose the data provenance and any known copyrighted content policies without revealing exact data points. This section enables downstream copyright infringement scans and data lineage verification.

A quantified disclosure of the computational resources used for training, typically measured in floating-point operations (FLOPs) or GPU-hours. It includes the hardware architecture, cloud provider, and estimated energy consumption in MWh. This data is critical for sustainable AI reporting and determining if a model crosses the systemic risk threshold of 10^25 FLOPs.

Standardized evaluation scores on public and internal benchmarks measuring reasoning, coding, multilingual performance, and factual accuracy. This section must include the model's hallucination rate benchmark and grounding score. It provides deployers with a baseline for model risk tiering and identifies potential dangerous capability benchmarks relevant to CBRN (Chemical, Biological, Radiological, Nuclear) risks.

A summary of the internal and external red-teaming report findings, detailing the model's jailbreak susceptibility and safety alignment threshold. It outlines the guardrail configuration used during testing, including Reinforcement Learning from Human Feedback (RLHF) methodologies. This component is essential for completing a downstream algorithmic impact assessment.

A precise definition of the model's designed purpose, supported modalities, and target deployment contexts. Crucially, it enumerates prohibited use cases and known limitations, such as high-risk classification scenarios without human oversight. This section directly informs the conformity assessment process and helps deployers avoid high-risk classification violations under the EU AI Act.

A structured record of the model's algorithmic supply chain, including the developer identity, release date, version identifier, and upstream dependencies. It details the model deprecation policy and API stability commitment. This section functions as a lightweight AI Bill of Materials (AIBOM), enabling procurement teams to assess vendor lock-in risk and verify interoperability standards.

A structured transparency document detailing a machine learning model's intended use, performance benchmarks, and limitations. Unlike the broader transparency report, a model card focuses on a single model's technical specifications, including evaluation results across different demographic groups and known failure modes. It serves as the primary unit of disclosure for model developers and is often a required component of a comprehensive transparency report.

A transparency artifact that documents the safety evaluation and operational context of an entire AI system, not just the model. It expands beyond the model card to include the user interface, deployment environment, and human oversight mechanisms. System cards are critical for understanding how a foundation model behaves when integrated into a downstream application, capturing risks that emerge from system-level interactions.

A formal, structured inventory of all software, data, and model components used to construct an AI system. Analogous to a software bill of materials (SBOM), the AIBOM provides a machine-readable manifest that enables downstream users to rapidly identify vulnerabilities, licensing conflicts, and supply chain dependencies. It is a foundational input for generating a comprehensive transparency report.

The documented end-to-end origin, movement, and transformation history of all datasets used to train a model. This includes data sources, collection methodologies, preprocessing steps, and any augmentation or filtering applied. Complete lineage is a mandatory disclosure element in transparency reports, enabling auditors to assess copyright compliance, bias introduction, and data quality.

A set of regulatory requirements specifically imposed on foundation models with broad applicability under the EU AI Act. These obligations mandate transparency reporting on training data, compute resources, and known capabilities. Providers must also cooperate with downstream deployers and disclose any systemic risk thresholds that their models exceed, making the transparency report a legal compliance artifact.

A regulatory mandate requiring developers to report to authorities when training runs exceed a specified computational power limit, typically measured in floating-point operations (FLOPs). This threshold serves as a proxy for capability and triggers heightened transparency obligations. The notification is a critical data point within the transparency report, signaling that a model may possess emergent or dangerous capabilities.