QomSSLab/Anonymizer-4b

QomSSLab/Anonymizer-4b is a fine-tuned Gemma 3 4B model designed to anonymize Persian legal texts by masking or replacing all personally identifiable information (PII). It is trained on the QomSSLab/Anonymized_Cases dataset.

💡 Use Cases

• Data privacy for legal document processing.
• Preprocessing step for building publicly shareable Persian legal corpora.
• Protecting PII in judicial NLP pipelines.

🧠 Model Details

• Base Model: Gemma 3 4B
• Language: Persian (Farsi)
• Training Data: Synthetic and real anonymized Persian legal cases.
• Task: Text-to-text generation (anonymization)

📦 Example Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model = AutoModelForCausalLM.from_pretrained("QomSSLab/Anonymizer-4b",  device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("QomSSLab/Anonymizer-4b")
tokenizer.add_eos_token = False

messages = [
    {"role": "system", "content": "You are a data privacy expert. Your task is to anonymize the following case text by removing or replacing all personally identifiable information (PII)."},
    {"role": "user", "content": "پرونده‌ای درباره ازدواج بین هانیه و عبدالرحیم با اطلاعات هویتی متعدد..."}
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, add_special_tokens=False)
inputs = tokenizer([prompt], return_tensors="pt", add_special_tokens=False).to("cuda")

outputs = model.generate(
    **inputs,
    max_new_tokens=400,
    temperature=0.1,
    top_p=0.95,
    top_k=64,
    disable_compile=True
)

anonymized_text = tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True)
print(anonymized_text)

📊 Evaluation

The model was evaluated qualitatively on a diverse collection of Persian legal documents. It effectively identifies and anonymizes a range of personally identifiable information (PII), including:

• Full names
• National IDs
• Addresses
• Dates of birth
• Case numbers
• Geographic locations

The model is particularly well-suited for preprocessing court cases for research, public data release, or downstream tasks like summarization and classification while preserving privacy.

Limitations

• May occasionally miss rare or out-of-distribution PII formats.
• Not guaranteed to anonymize very short or extremely noisy texts.
• Trained primarily on formal legal language; performance may degrade on informal Persian.

📁 Dataset

This model was fine-tuned on the QomSSLab/Anonymized_Cases dataset, which includes manually and synthetically anonymized court documents and legal filings in Persian. The dataset contains a mix of real and simulated entities, helping the model generalize across varied legal formats and writing styles.