Distil-PII-SmolLM2-135M-Instruct

A small language model (SLM) fine-tuned by Distil Labs for policy-aware PII redaction that outputs a single JSON object with redacted_text and entities. Optimized to run locally with strong accuracy and strict schema adherence.

Model Details

• Developed by: Distil Labs GmbH
• License: Apache 2
• Finetuned from: HuggingFaceTB/SmolLM2-135M-Instruct

Intended Use & Limitations

• Use cases: Redacting support chats, logs, tickets, transcripts—removing identity while preserving ops signals (IDs last-4, order numbers, etc.).
• Out of scope: Legal or compliance advice; languages beyond English (generalization not guaranteed); domain-specific IDs unseen in training.

Input & Output

Input: A plain-text prompt with task instruction + context. Output (JSON only):

{
  "redacted_text": "Text with in-place tokens",
  "entities": [
    {"value": "<original>", "replacement_token": "[TOKEN]", "reason": "<why>"}
  ]
}

Tokens: [PERSON] [EMAIL] [PHONE] [ADDRESS] [SSN] [ID] [UUID] [CARD_LAST4:####] [IBAN_LAST4:####] [GENDER] [AGE] [RACE] [MARITAL_STATUS]

Training

Instruction-tuned on a compact policy spec + ~20 curated examples emphasizing exact JSON schema, minimal in-place edits, and entity correctness.

Evaluation

Judged by a frontier LLM using a deterministic rubric: JSON-only, schema validity, redacted_text exact match, and set-equality of (value, replacement_token) pairs (reason/order ignored). Score: 0.25 +/- 0.05.

How to Use

Details of deployment can be found in https://docs.distillabs.ai/how-to/model-deployment

Risks & Mitigations

• False negatives/positives: May miss novel formats or over-redact generic terms. Mitigate via guardrails + post-validation.
• Policy drift: Keep task preamble fixed; monitor with unit tests.

Model Sources

• Homepage: https://distillabs.ai
• Contact: [email protected]