---
library_name: transformers
license: apache-2.0
language:
  - zh
  - en
pipeline_tag: image-text-to-text
tags:
  - forgery-detection
  - document-forensics
  - image-tampering
  - vision-language-model
  - vlm
  - qwen3.5-vl
---

<p align="center">
  <img src="docshield_showcase.png" alt="DocShield-9B showcase" width="70%">
</p>

# DocShield-9B

**DocShield-9B** is a forensic-grade vision-language model for **document / image forgery analysis**. It inspects an input document image, reasons over visual tampering traces and logical consistency, and produces a structured forgery-analysis report with localized tampered regions (grounding coordinates), per-region reasoning, an overall conclusion, and a fraud-risk score.

It is fine-tuned from **Qwen3.5-VL-9B** with supervised training on forensic document-forgery data, and supports Qwen3.5 thinking mode.

📄 **Paper:** [arxiv.org/abs/2604.02694](https://arxiv.org/abs/2604.02694)

## Training data

DocShield-9B was developed using the **RealText** forensic document datasets:

- [vankey/RealText-V1](https://huggingface.co/datasets/vankey/RealText-V1)
- [vankey/RealText-V2](https://huggingface.co/datasets/vankey/RealText-V2)

## Capabilities

- **Visual forgery trace analysis** — crude redaction / mosaic, font & anti-aliasing inconsistency, edge halos, copy-paste artifacts, compression mismatches.
- **Logical & fact-checking** — price/quantity contradictions, date conflicts, bulk-discount logic violations.
- **Semantic alteration detection** — subtle spec substitutions (e.g. color, material) that bypass crude visual checks.
- **Localization** — bounding-box coordinates for each tampered region with per-anomaly reasoning.
- **Structured report** — conclusion (`FORGED` / authentic) + fraud-risk score.

## Model details

| | |
|---|---|
| Base model | Qwen3.5-VL-9B |
| Architecture | Qwen3_5ForConditionalGeneration (hybrid linear / full attention) |
| Precision (weights) | bfloat16 |
| Max new tokens | 1024 (default) |
| Thinking mode | supported (`--thinking` / `--no-thinking`) |

> The base model is **not** bundled here. This repository only releases the
> fine-tuned DocShield-9B weights.

## Quick start

Install dependencies:

```bash
pip install -U transformers torch torchvision pillow qwen-vl-utils
```

> Requires a `transformers` version with native **Qwen3.5-VL (`qwen3_5`)** support.
> The released weights were saved with `transformers==5.13.0`.

Run inference (see `inference.py` in this repo):

```bash
# default: greedy, thinking disabled
python inference.py --image path/to/image.jpg --no-thinking --max-new-tokens 1024

# with sampling + thinking mode
python inference.py --image path/to/image.jpg --thinking --do-sample \
                    --temperature 0.6 --top-p 0.8 --top-k 20 --max-new-tokens 2048
```

### Minimal example

```python
import torch
from transformers import AutoProcessor, AutoTokenizer, AutoModelForImageTextToText
from qwen_vl_utils import process_vision_info

MODEL = "vankey/DocShield-9B"

tokenizer = AutoTokenizer.from_pretrained(MODEL, trust_remote_code=True)
processor = AutoProcessor.from_pretrained(MODEL, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
    MODEL, torch_dtype=torch.bfloat16, trust_remote_code=True, device_map="auto"
)
model.eval()

SYSTEM_PROMPT = (
    "你是一个图像鉴伪专家，擅长结合视觉，文字结合伪造特征分析手段鉴别输入图像的真假。"
    "分析过程中，你会逐步分析，抽丝剥茧，找到图像伪造的蛛丝马迹，最终给出专业的鉴别结果及分析。"
)
USER_PROMPT = "请分析这张文档图片是否存在伪造或篡改风险，并输出一份专业、精炼、准确的防伪分析报告。"

messages = [
    {"role": "system", "content": [{"type": "text", "text": SYSTEM_PROMPT}]},
    {"role": "user", "content": [
        {"type": "image", "image": "image.jpg"},
        {"type": "text", "text": USER_PROMPT},
    ]},
]

text = processor.apply_chat_template(messages, tokenize=False,
                                     add_generation_prompt=True, enable_thinking=False)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(text=[text], images=image_inputs, videos=video_inputs,
                   padding=True, return_tensors="pt").to(model.device)

with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=1024, do_sample=False)

gen = [o[len(i):] for i, o in zip(inputs["input_ids"], out)]
print(processor.batch_decode(gen, skip_special_tokens=False)[0])
```

## Inference notes

- **Image input** — pass the image path directly in the message content; the processor handles resize/tokenization.
- **Thinking mode** — `--no-thinking` (default) gives a direct report; `--thinking` enables Qwen3.5 chain-of-thought before the report (use a larger `--max-new-tokens`).
- **Decoding** — greedy by default (`do_sample=False`); pass `--do-sample` with `--temperature/--top-p/--top-k` for sampling.
- **Precision** — `bfloat16` is the tested configuration (`--dtype bf16`).

## Citation

```bibtex
@article{docshield2026,
  title={DocShield: A Forensic Vision-Language Model for Document Forgery Analysis},
  author={DocShield},
  year={2026},
  url={https://arxiv.org/abs/2604.02694}
}
```

## License

Apache-2.0.