AXONVERTEX-AI-RESEARCH/qwen2.5-0.5b-pii-anonymizer

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jun 12, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

The AXONVERTEX-AI-RESEARCH/qwen2.5-0.5b-pii-anonymizer is a 0.5 billion parameter Qwen2.5-based causal language model fine-tuned for context-aware PII anonymization. It replaces personally identifiable information with realistic synthetic data, preserving text grammar, formatting, and semantic flow. This model excels at generating natural-sounding anonymized text across 55+ PII categories, making it suitable for data privacy and compliance applications.

Loading preview...

Overview

This model, qwen2.5-0.5b-pii-anonymizer, is a 0.5 billion parameter Qwen2.5-based instruction-tuned model developed by AXONVERTEX-AI-RESEARCH. Its core function is context-aware PII anonymization, replacing sensitive data with realistic synthetic information while maintaining the original text's integrity. Unlike traditional redaction that uses placeholders, this model generates natural-sounding replacements, crucial for downstream NLP tasks and data utility.

Key Capabilities

  • Realistic PII Replacement: Substitutes PII (names, emails, phone numbers, addresses, SSNs, etc.) with synthetic data that preserves grammar, formatting, tone, and semantic flow.
  • Broad PII Coverage: Handles over 55 PII categories across corporate, medical, financial, conversational, and multilingual contexts.
  • Context-Awareness: Identifies PII that rule-based systems might miss and generates contextually appropriate replacements.
  • Efficient Fine-tuning: Trained using QLoRA (4-bit NF4 quantization + LoRA r=16) on the Qwen2.5-0.5B-Instruct base model, achieving convergence in approximately 2.3 hours.
  • Multilingual Support: While primarily English-focused, it includes training data for French, German, and Italian, offering some multilingual capabilities.

Use Cases

This model is ideal for scenarios requiring data privacy without sacrificing data utility, such as:

  • Anonymizing customer support tickets, medical records, or financial documents for analytics or sharing.
  • Preparing datasets for training other models where PII needs to be removed but context preserved.
  • Developing privacy-preserving applications where natural language flow is critical.

Limitations

As a Phase 1 model, it has limitations including occasional artifact tokens, potential over-replacement of non-PII context, and inconsistent multilingual performance. It is not intended as a standalone compliance mechanism and is designed to be integrated into a broader PII anonymization pipeline (e.g., with Microsoft Presidio and GLiNER).