vankey/DocShield-7B

<p align="center"> <img src="docshield_showcase.png" alt="DocShield-7B showcase" width="90%"> </p>

DocShield-7B

DocShield-7B is a forensic-grade vision-language model for document / image forgery analysis. It inspects an input image, reasons step-by-step over visual tampering traces and logical consistency, and produces a professional forgery-analysis report with localized tampered regions and a forgery score.

It is fine-tuned from Qwen2.5-VL-7B-Instruct with supervised chain-of-thought (CoT) reasoning on document-forgery data.

📄 Paper: arxiv.org/abs/2604.02694

Capabilities

• —Visual forgery trace analysis — font / glyph / kerning / baseline inconsistency, copy-paste artifacts, edge halos, compression mismatches, noise-pattern breaks.
• —Logical & fact-checking — impossible dates, failed calculations, contradictory metadata, domain-commonsense violations.
• —Localization — bounding boxes of tampered regions with per-region reasoning.
• —Structured CoT report — forensic-style report with a final conclusion and forgery score.

Model details

The weights are stored in bfloat16 (~16 GB). Always load them with torch_dtype=torch.float32 so computation runs in float32 (the bf16 weights are upcast on load). The base model is not bundled here — download it from Qwen/Qwen2.5-VL-7B-Instruct if needed. This repository only releases the fine-tuned DocShield-7B weights.

Quick start

Install dependencies:

pip install transformers torch torchvision pillow opencv-python qwen-vl-utils

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

python inference.py --image path/to/image.jpg

Minimal example

import cv2, torch
from PIL import Image
from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration
from qwen_vl_utils import process_vision_info

MODEL_PATH = "vankey/DocShield-7B"

model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    MODEL_PATH,
    torch_dtype=torch.float32,
    attn_implementation="eager",
    device_map="auto",
)
model.eval()
processor = AutoProcessor.from_pretrained(MODEL_PATH)

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 = "请帮我分析这张图片是否是伪造的,并给出分析报告."

image = cv2.cvtColor(cv2.resize(cv2.imread("image.jpg"), (1344, 896)), cv2.COLOR_BGR2RGB)
image = Image.fromarray(image)

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

text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
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,
                         do_sample=True, temperature=1.0, top_p=1.0,
                         top_k=0, repetition_penalty=1.0,
                         max_new_tokens=8192)

generated = out[:, inputs["input_ids"].shape[1]:]
print(processor.batch_decode(generated, skip_special_tokens=True)[0])

Inference notes (important)

These choices are required to reproduce the reported results:

1. 1.Image resize — resize the input to 1344 × 896 (W × H) before processing: cv2.resize(image, (1344, 896)).
2. 2.Precision — load with `float32`, never compute in `bfloat16`. The weights are stored as bfloat16; load them with torch_dtype=torch.float32 so they are upcast and computation runs in float32. Computing in bfloat16 accumulates rounding error over the long CoT and degrades output into gibberish on harder images. float16 + eager attention also overflows (NaN) for long contexts. float32 (compute) + eager is the verified configuration.
3. 3.Sampling — `temperature=1.0, top_p=1.0, top_k=0, repetition_penalty=1.0` (full multinomial sampling). Do not use the values in the bundled generation_config.json (temperature=0.1, top_k=1, top_p=0.001, repetition_penalty=1.05) — that near-greedy config triggers repetition loops.
4. 4.No flash-attention — attn_implementation="eager".

Citation

@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.