Skip to main content

Ant Opens Up About AI Training Data to Meet EU Rules

Ant's inclusionAI Pulls Back the Curtain on AI Training Data

Ant Financial's inclusionAI has launched a new repository on HuggingFace called AI-Transparency. But don't expect to download model weights or datasets here. Instead, they're publishing summaries of what went into training their Ling, Ring, and Ming series models. It's a deliberate move to answer a question that's often left in the dark: what data did the model actually learn from?

The repository covers a wide range of models. On the Ling side, you've got everything from Ling-2.0 up to Ling-3.0, including the LLM versions (flash, tiny, flash-fin) and the VL (visual language) variant, Ling-3.0-flash-VL. The Ring series spans 2.0, 2.5-1T, and 2.6-1T. And the Ming family includes the Omni multimodal models (like flash-omni-preview and flash-omni-2.0) plus two unified audio and visual models: UniAudio-16B-A3B and UniVision-16B-A3B.

Each summary sticks strictly to the model and version it's about. If a document mentions Ling-2.0, it only covers Ling-2.0 — including Ling1T, Ling-flash2.0, and Ling-mini-2.0 — and doesn't spill over into other models or quantized versions. The documents are all versioned and dated, and they describe training content at a high level without handing over the actual training data or granting any rights to third parties.

Why This Matters: The EU AI Act Connection

This isn't just a random act of openness. The framework follows the public training content summary template tied to Article 53(1)(d) of the EU AI Act (Regulation 2024/1689). In plain English: this is a standardized format that general-purpose AI model providers in the EU are expected to use. So Ant is essentially checking a compliance box — but doing it in a way that also gives outsiders a peek under the hood.

It's worth noting that the official page stresses these documents aren't an endorsement or certification. They don't mean every inclusionAI model is covered. And the dates in the PDFs are just document version dates, not when the models first went live. For real changes, you'd need to check the commit history in the repository.

A Small Window into Model Development

As global regulation turns "training transparency" from a nice-to-have into a hard requirement, Ant's systematic release of these summaries serves two purposes. First, it paves the way for overseas compliance. Second, it quietly opens a window into how their models were developed — something that's usually kept behind closed doors.

The contact page points to Ant's official developer site for Ling. And the version history reminds readers that the PDF dates are just document versions, not model launch dates. For actual changes, check the repository's commit records.

Key Points:

  • AI-Transparency repository on HuggingFace shares training data summaries, not weights or datasets.
  • Covers Ling, Ring, and Ming model series, with strict version scoping.
  • Follows EU AI Act Article 53(1)(d) template for public training content summaries.
  • Documents are versioned and dated, but don't grant rights to training data.
  • Ant emphasizes this isn't certification or endorsement — just compliance and a peek into development.