Distil-PII: family of PII redaction SLMs (github.com)

Distil Labs released Distil-PII, a family of small language models (SLMs) fine-tuned for policy-aware PII redaction and packaged for on-device or on-prem deployment (gguf builds for Llama-3.2 3B/1B, gemma-3 270M, and SmolLM2 135M). After targeted distillation and task-specific fine-tuning, the 1B student achieves an LLM-as-judge score of 0.81 ± 0.02—virtually matching a 600B+ baseline (DeepSeek 3.1 at 0.84 ± 0.03)—while offering much lower latency, cost, and no data egress. This makes practical large-scale redaction of logs, transcripts, and support data feasible where privacy, throughput, and strict schema conformance matter. Technically, models are trained to output a single JSON with minimal in-place redactions (redacted_text) and an entities array of {value, replacement_token, reason}. They redact PERSON, EMAIL, PHONE, ADDRESS, SSN, ID, UUID, CREDIT_CARD (keep last-4 as [CARD_LAST4:1234]), IBAN (keep last-4), GENDER, AGE, RACE, MARITAL_STATUS, and preserve operational tokens. Distillation addressed common SLM failure modes—schema drift, hallucinated entities, under/over-redaction—by generating and validating synthetic in-domain data and finetuning students to emulate a teacher. Evaluation uses DeepSeek 3.1 as a deterministic judge checking JSON-only output, schema validity, exact redacted_text match, and set-equality of entity (value, replacement_token) pairs, prioritizing policy adherence over style. The result: high-performing, deployment-ready SLMs that enable safe data use without heavyweight LLM infrastructure.