Models
The models that power the product.
Search, Agent, Verify, and Redact are only as good as the models underneath them. Obiter treats those models as product: measured, versioned, and something we keep improving. Our redaction models are being released as open source.
Capabilities
Open redaction
Rampart is open source, and our redaction models are being released as open source, so you can inspect what reviews your documents.
Rampart today
The live redaction detector, plus a UK legal supplement.
Legal fine-tunes
Tuned on legal text whose licence permits it, evaluated before promotion.
Our own models
Obiter-trained models, starting with redaction.
Redaction runs on an open-source detector today. Fine-tunes and our own trained models are how we will improve the rest of the system.
Why models are a product surface
The models are product: measured, versioned, and improving.
- What runs today, what we fine-tune, what we train
- Open-source redaction models you can inspect
- The workspace inherits whatever limits the models have
Redaction model today
Redact runs on Rampart, an open-source privacy detector from National Design Studio, with a UK legal supplement on top.
- Open redaction detector in the live Redact path
- UK supplement for legal-specific patterns
- Versioned — a run says which detector produced it
- Reviewable spans, not silent automatic scrubbing
Fine-tuned models
Fine-tuning on legal text whose licence permits it, not generic web text.
- Legal text we hold the rights to use, and real label spaces
- Pinned revisions, evaluated before promotion
- Never trained by mixing client data across organisations
Building our own
Obiter intends to train models of its own.
- Retrieval, agent assistance, verification, redaction
- A stack we can measure and improve
- No permanent dependence on a closed frontier API
What is live, and what is next
Live today: the open detector behind Redact. Next: fine-tunes, then Obiter-trained models.
- Legal fine-tunes of the detector come first
- Further models as Agent and Verify land
- Redaction models released as open source
Questions
Which model runs redaction today?
Rampart, an open-source privacy detector, with our UK legal supplement on top.
Are the models open source?
Rampart is open source, and we are releasing our redaction models as open source. Other models are not covered by that statement.
Is client data used for training?
Never. Client matter data is not mixed into training across organisations.
Work with Models inside Obiter.
In development. Request access to hear when it reaches the workspace.