Commitment

Safety Data Transparency Commitment

AI alignment is a shared human problem. Safety training data should not disappear into private procurement channels, undisclosed licensing conversations, or unverifiable claims of quality.

1. Commitment to discoverability

The signer commits to making safety-relevant training datasets discoverable through a public listing, registry, marketplace, paper, model card, or equivalent public reference.

2. Commitment to evidence

The signer commits to publishing or sharing enough information for buyers, researchers, and evaluators to understand the dataset's intended safety use case, collection method, known limitations, and evaluation evidence.

3. Commitment to access clarity

The signer does not need to make safety data free. The signer does commit to making access terms clear, including whether the dataset is open, commercial, research-access, restricted, or unavailable.

4. Commitment to benchmark relevance

Where possible, the signer will describe how the dataset relates to safety and alignment benchmarks, including measured lift, model evaluations, or known gaps.

5. Commitment to responsible disclosure

The signer may withhold details that would create material misuse risk, violate privacy, breach law, or expose sensitive security procedures. Those limits should be named rather than hidden.

6. Commitment to update stale information

The signer will make a reasonable effort to update public dataset information when access terms, benchmark evidence, or safety limitations materially change.

Signature fields

  • Organization
  • Signer name
  • Signer title
  • Contact email
  • Dataset or program covered
  • Signature date

Sign the commitment

Complete the form below to start the signing process. We will follow up by email.

Start signing process

This commitment is a public transparency pledge, not legal advice. Organizations should review it with counsel before signing.