---
base_model: Qwen/Qwen2.5-1.5B-Instruct
library_name: peft
license: apache-2.0
language:
- en
pipeline_tag: text-generation
---

# Adaptively-tuned Qwen2.5-1.5B Paraphraser

This model is an adaptively fine-tuned version of Qwen2.5-1.5B-Instruct optimized to evade the EXP watermarking method while preserving text quality. It serves as a paraphrasing model that maintains semantic meaning while modifying the statistical patterns used for watermark detection.

## Model Details

### Model Description

This model is a fine-tuned version of Qwen2.5-1.5B-Instruct that has been optimized using Direct Preference Optimization (DPO) to evade the [EXP watermarking method](https://www.scottaaronson.com/talks/watermark.ppt) described in Aaronson and Kirchner (2023). The model preserves text quality while modifying the statistical patterns that watermarking methods rely on for detection.

- **Model type:** Decoder-only transformer language model
- **Language(s):** English
- **Finetuned from model:** Qwen2.5-1.5B-Instruct

## Get Started

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

# Load the base model
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")

# Load the LoRA adapter
model = PeftModel.from_pretrained(model, "DDiaa/WM-Removal-EXP-Qwen2.5-1.5B")

# Prepare the prompt

system_prompt = (
    "You are an expert copy-editor. Please rewrite the following text in your own voice and paraphrase all "
    "sentences.\n Ensure that the final output contains the same information as the original text and has "
    "roughly the same length.\n Do not leave out any important details when rewriting in your own voice. Do "
    "not include any information that is not present in the original text. Do not respond with a greeting or "
    "any other extraneous information. Skip the preamble. Just rewrite the text directly."
)

def paraphrase_text(text):
    # Prepare prompt
    prompt = tokenizer.apply_chat_template(
        [
            {"role": "system", "content": system_prompt},
            {"role": "user", "content": f"\n[[START OF TEXT]]\n{text}\n[[END OF TEXT]]"},
        ],
        tokenize=False,
        add_generation_prompt=True,
    ) + "[[START OF PARAPHRASE]]\n"
    
    # Generate paraphrase
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    outputs = model.generate(
        **inputs,
        max_new_tokens=512,
        temperature=1.0,
        do_sample=True,
        pad_token_id=tokenizer.pad_token_id
    )
    
    # Post-process output
    paraphrased = tokenizer.decode(outputs[0], skip_special_tokens=True)
    paraphrased = paraphrased.split("[[START OF PARAPHRASE]]")[1].split("[[END OF")[0].strip()
    
    return paraphrased
```


## Uses

### Direct Use

The model is designed for research purposes to:
1. Study the robustness of watermarking methods
2. Evaluate the effectiveness of adaptive attacks against content watermarks
3. Test and develop improved watermarking techniques

### Downstream Use

The model can be integrated into:
- Watermark robustness evaluation pipelines
- Research frameworks studying language model security
- Benchmark suites for watermarking methods

### Out-of-Scope Use

This model should not be used for:
- Production environments requiring watermark compliance
- Generating deceptive or misleading content
- Evading legitimate content attribution systems
- Any malicious purposes that could harm individuals or society

## Bias, Risks, and Limitations

- The model inherits biases from the base Qwen2.5-1.5B-Instruct model
- Performance varies based on text length and complexity
- Evasion capabilities may be reduced against newer watermarking methods
- May occasionally produce lower quality outputs compared to the base model
- Limited to English language texts

### Recommendations

- Use only for research and evaluation purposes
- Always maintain proper content attribution
- Monitor output quality metrics
- Consider ethical implications when studying security measures
- Use in conjunction with other evaluation methods



## Citation

**BibTeX:**
```bibtex
@article{diaa2024optimizing,
  title={Optimizing adaptive attacks against content watermarks for language models},
  author={Diaa, Abdulrahman and Aremu, Toluwani and Lukas, Nils},
  journal={arXiv preprint arXiv:2410.02440},
  year={2024}
}
```

## Model Card Contact

For questions about this model, please file an issue on the GitHub repository: https://github.com/ML-Watermarking/ada-llm-wm