What data is shown?

The table is generated from the public-ai sovereignty pipeline, which evaluates models using Hugging Face metadata and, optionally, supporting web evidence. Each entry includes a sovereignty score and a breakdown across key dimensions.

In this demo, the dataset is a static snapshot (data/models.json). In a full deployment, this would be continuously generated by analysing live model cards, organisation metadata, and external documentation.

What “sovereignty” means here

In this dashboard, sovereignty refers to the degree of control, independence, and local alignment of an AI model. This includes:

• Whether the model can be run and controlled independently
• Whether its development is tied to a specific country or institution
• How transparent its data and training process are

This definition intentionally combines technical, organisational, and geopolitical factors into a single comparative framework.

Dimensions used

Each model is evaluated across six dimensions:

• Training data privacy – whether the dataset is private or openly described
• Local training – whether the model is developed within a specific national or institutional context
• Control of weights – whether weights are open and can be self-hosted
• Model independence – whether the model is trained from scratch or derived from another model
• Weight privacy – whether weights are restricted or publicly available
• Country-specific knowledge – whether the model reflects local language, policy, or culture

Each dimension is scored between 0 and 1, where higher values indicate greater sovereignty along that axis.

How scores are calculated

Each dimension starts with a neutral score of 0.5, then is adjusted using heuristics derived from model metadata:

• Open licenses increase control of weights but reduce weight privacy
• Presence of a base model lowers independence (indicating fine-tuning)
• Known public or state-backed organisations increase local and country-specific scores
• Transparency signals (e.g. “open”, “transparent”) reduce training data privacy

When web evidence is enabled, the system can either:

• Apply keyword-based adjustments (e.g. “sovereign”, “open data”)
• Or use a language model to estimate scores from external sources

The final score is computed as a weighted average of all six dimensions (equal weighting by default), and scaled to a 0–100 range.

Data sources and signals

The pipeline relies primarily on structured metadata from Hugging Face, including model cards, tags, licences, and organisation names. Additional signals may be extracted from external web documents when available.

Because metadata is often incomplete or inconsistent, the system uses heuristic pattern matching rather than strict verification. This allows broad coverage but introduces some uncertainty.

Interpretation and limitations

This score is a heuristic approximation, not a definitive label. It blends multiple interpretations of sovereignty—such as openness, control, and national alignment—which can sometimes conflict. For example, open weights increase usability and control, but reduce privacy.

Scores are sensitive to metadata quality and should be interpreted as directional indicators rather than precise measurements. The dashboard is best used for comparison and exploration, alongside detailed model documentation.