---
library_name: peft
base_model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
license: apache-2.0
pipeline_tag: text-generation
language:
  - en
tags:
  - lora
  - peft
  - anonymization
  - pii
  - privacy
  - gdpr
  - knowledge-distillation
---

# TinyLlama LoRA — Reversible PII Anonymizer

LoRA adapters that turn TinyLlama-1.1B into a **local privacy filter**: it removes
personally identifiable information from text and returns a structured JSON object
containing both the anonymized text and a full mapping of every replaced token.

Because the mapping is returned, the process is **reversible**. Sanitized text can be
sent to an external LLM API; when the response comes back, the original values are
restored locally. Raw PII never leaves your infrastructure.

**Live demo:** https://huggingface.co/spaces/LorenzoMascia/tinyllama-lora-anonymizer
**Training pipeline:** https://github.com/LorenzoMascia/llm-distillery
**Write-up:** https://www.howai.cloud/fine-tuning-knowledge-distillation-llm.html

## Example

Input:

```
Mario Rossi lives at Via Roma 1, Milano. Email: m.rossi@email.com
```

Output:

```json
{
  "anonymized_text": "[NAME_1] lives at [ADDRESS_1]. Email: [EMAIL_1]",
  "replaced_tokens": [
    {"replaced_value": "[NAME_1]",    "original_value": "Mario Rossi"},
    {"replaced_value": "[ADDRESS_1]", "original_value": "Via Roma 1, Milano"},
    {"replaced_value": "[EMAIL_1]",   "original_value": "m.rossi@email.com"}
  ]
}
```

## Usage

These are LoRA adapters — the base model weights are **not** included.

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

BASE = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
tokenizer = AutoTokenizer.from_pretrained(BASE)
model = AutoModelForCausalLM.from_pretrained(BASE, torch_dtype=torch.float16, device_map="auto")
model = PeftModel.from_pretrained(model, "LorenzoMascia/tinyllama-lora-anonymizer")
```

### Prompt format (required)

The adapters were trained on the Alpaca template. Using a different format will
degrade output quality significantly.

```
### Instruction:
Analyze the following text and anonymize all personally identifiable information (PII). Return a JSON object with the anonymized text and all replaced tokens.

### Input:
{your text here}

### Response:
```

Generate with a low temperature (~0.1) — the output is a data structure, not prose.

## Training

- **Method:** LoRA (PEFT) — rank 16, target modules `q_proj`, `v_proj`
- **Base model:** TinyLlama-1.1B-Chat-v1.0 (1.1B parameters)
- **Data:** 100% synthetic, generated by a GPT-4 teacher model via the OpenAI API.
  No real personal data was collected or used at any stage.

Training an anonymizer would normally require a corpus of real PII — precisely the
data the model is meant to protect. Synthetic generation resolves that: the teacher
produces realistic input/output pairs, and the student learns from those alone.

## Limitations

- **Not formally evaluated.** No benchmark numbers are published. Treat this as a
  reference implementation, not a validated production component.
- **False negatives are the risk that matters.** A missed entity is data that leaves
  your perimeter. In production, pair this with rule-based detection for
  deterministic formats (email, IBAN, credit card, VAT) and validate before sending.
- **Teacher distribution bias.** The model learned PII as GPT-4 imagines it, not as
  it appears in messy real-world documents (OCR artifacts, abbreviations, mixed
  languages).
- **Language coverage.** Primarily English; behaviour on other languages is untested.
- **Structured output can fail.** A 1.1B model occasionally emits malformed JSON.
  Parse defensively, or use grammar-constrained decoding.

## License

Apache 2.0