ganmoor-ai-labs/piiguard-qwen3-1.7b

PIIGuard — Indian-language PII extraction (Qwen3-1.7B)

PIIGuard extracts personally identifiable information from English, Hindi (Devanagari), and Hinglish (code-mixed) enterprise text and returns structured JSON — built for on-prem redaction pipelines where data cannot leave the building (support transcripts, logs, KYC notes, LLM-gateway scrubbing, DPDP compliance).

It handles the India-specific surface forms Western PII tools miss: Aadhaar numbers (including Devanagari digits ९८७६…), PAN codes, UPI/VPA handles, IFSC codes, Indian address formats — while deliberately ignoring lookalike traps (12-digit UTR references, order/ticket/invoice numbers, OTPs, GSTINs, employee IDs, company names).

The Q4KM GGUF is 1.2GB and runs on CPU-only hardware down to a Raspberry Pi via llama.cpp/Ollama.

Entity types

PERSON PHONE EMAIL AADHAAR PAN ADDRESS DOB BANK_ACCOUNT IFSC CARD UPI_ID

Explicitly not flagged (by design): GSTIN (public business identifier), UTR/transaction refs, OTPs, order/ticket/invoice/tracking numbers, employee IDs, company names, standalone city names, non-birth dates.

Results

Held-out test set: 1,320 examples from unseen templates and unseen PII values (train/test value pools are hash-disjoint), 200 of them PII-free.

Per language F1: en 98.4 / hi 98.8 / hi-en 99.2. Full per-entity × per-language tables in `eval_report.md`.

Q4KM quantization spot-check (100-example slice): recall 98.4%, precision 95.9%.

Input/output contract

• —system: Extract all personally identifiable information (PII) from the user's text. Respond with only a JSON object: {"entities": [{"type": ..., "text": ...}]}. Valid types: PERSON, PHONE, EMAIL, AADHAAR, PAN, ADDRESS, DOB, BANK_ACCOUNT, IFSC, CARD, UPI_ID. If there is no PII, return {"entities": []}.
• —user: raw text
• —assistant: {"entities": [{"type": "PHONE", "text": "९८२२०४५६७१"}, ...]} — text is the verbatim substring; locate spans by string search downstream. Use temperature 0.

Quickstart (Ollama — easiest)

ollama run santosh07401/piiguard "mera number ९८२२०४५६७१ hai aur naam Ravi Gupta"
# {"entities":[{"type":"PHONE","text":"९८२२०४५६७१"},{"type":"PERSON","text":"Ravi Gupta"}]}

The system prompt is baked in — paste text, get JSON. Model page: ollama.com/santosh07401/piiguard

Quickstart (llama.cpp)

llama-server -m piiguard-q4_k_m.gguf --port 8090 -c 4096
curl -s localhost:8090/v1/chat/completions -d '{
  "messages": [
    {"role": "system", "content": "<system prompt above>"},
    {"role": "user", "content": "sir mera naam Anjali Deshmukh hai, mera number ९८२२०४५६७१ hai"}
  ],
  "temperature": 0, "max_tokens": 512
}'

Files

Training

LoRA (r=16, α=32) on Qwen/Qwen3-1.7B with TRL SFTTrainer, assistant-only loss, thinking disabled; 3 epochs + 1 targeted fix epoch on a single NVIDIA GB10 (DGX Spark). Data: ~26k fully synthetic examples — carrier templates written by gemma-3-27b-it (local), filled with programmatically generated, checksum-valid fake PII (Verhoeff-valid Aadhaar, Luhn-valid cards, region-correct PIN codes; Roman + Devanagari name pairs). Labels are exact by construction; no real personal data was used at any point.

Limitations

• —Trained on synthetic text in 3 language tracks; other Indic languages (Bengali, Tamil, Telugu…) are not yet covered.
• —Entity text is returned verbatim but span location is left to the caller.
• —16-digit bank accounts vs. cards are disambiguated by context; genuinely context-free digit runs can still be typed wrong.
• —Not a compliance guarantee: use as a component in a reviewed pipeline, with human escalation where stakes require it.
• —PASSPORT / VOTERID / DRIVINGLICENSE are not yet entity types (roadmap).