Ertas for PII Redaction Pipelines
Train a small model in Ertas Studio that finds and redacts names, emails, phone numbers, addresses and ID numbers in your text. Export it and run it on your own servers, so the documents it processes stay with you. For teams cleaning text before it goes into AI training sets, search indexes or downstream tools.
The Challenge
Organizations handling client data for AI/ML projects must redact PII before any model training or RAG ingestion. Manual redaction is slow and error-prone. Regex-based scripts miss edge cases. Cloud redaction tools require data egress that regulated clients prohibit.
The Solution
With Ertas Studio you can train a model that redacts text the way your data needs. Each training example pairs a passage of text (the USER turn) with the same passage rewritten so personal details are replaced by placeholders such as [NAME], [EMAIL] or [MEDICAL_ID] (the ASSISTANT turn). Build those examples in Data Craft, where reviewers edit, version and approve them, or upload them as JSONL. Because the model learns from examples, it can pick up the formats regex scripts miss, such as names in unusual positions, local ID formats or addresses split across several lines.
Training runs on Ertas's cloud GPUs, so the training set is uploaded to Ertas. Build it from synthetic records or text that has already been de-identified, and keep the raw documents you plan to redact out of it. Train several small models in parallel, check each against held-out passages for missed and over-redacted entities, then export the best one from Hub as GGUF or full safetensors. It runs on your own servers with llama.cpp, Ollama or vLLM, so the documents it redacts stay in your infrastructure. Treat its output like any other redaction step: review samples, and keep a person checking the cases that matter most.
Key Features
PII Rules Defined by Examples
Choose which entity types to redact and how (mask, replace with a typed placeholder, or remove) by showing the model examples of each. To change the rules, change the examples and retrain.
Review Examples in Data Craft
Build training pairs as USER and ASSISTANT turns, edit and version them, and track which ones a reviewer has approved, so the model learns from redactions your team has signed off.
Measure Misses Before Deployment
Train several runs in parallel and compare them on held-out passages, counting missed entities and over-redacted text, before you pick the model to deploy.
Runs on Your Own Servers
Export the model as GGUF or full safetensors and run it with llama.cpp, Ollama or vLLM inside your network. The documents you redact at inference time stay there.
Example Workflow
A service provider receives client healthcare documents for a clinical NLP project and needs patient names, medical record numbers and addresses removed before anyone on the project reads them. The team generates a few thousand synthetic clinical notes in the client's format, with fictional patients, record numbers and addresses, and builds training pairs in Data Craft: the note as the USER turn and the redacted note as the ASSISTANT turn. The client's privacy lead reviews and approves a sample of the examples. In Ertas Studio the team trains runs on Qwen 2.5 1.5B and Llama 3.2 3B in parallel, and checks each against held-out synthetic notes plus a small set the client de-identified by hand. They export the better model from Hub as GGUF, and the client runs it with llama.cpp on a server inside its own network, over the real documents. The client's staff spot-check the redacted output before it is released to the NLP team.
Compliance & Security
A redaction model is one control in a de-identification process. Whether that process meets GDPR data minimisation, the HIPAA de-identification standard or the EU AI Act's data governance requirements is for your organisation and its advisers to decide. What Ertas changes is where the work happens. Once exported, the model runs on your own servers, so the documents it processes stay in your infrastructure. Training runs on Ertas's cloud GPUs, so keep the training set synthetic or already de-identified, and delete datasets from Studio when you no longer need them. Review a sample of redacted output before relying on it, since no model catches every entity.
Related Resources
Ship AI that runs on your users' devices.
Start free with 150 credits after card verification. Paid plans from $20/mo USD.