SEED

Simple ViT and Evolving Harness for Explainable Text Forgery Detection

---

🏆 3rd Place Solution for the ACM MM 2026 GenText-Forensics Challenge — detecting, localizing, and explaining text-centric document forgeries.

---

🧠 Overview

SEED is a modular forgery analysis pipeline with three stages:

Stage	Component	Description
1️⃣	Synthetic Data	Similarity-guided forgery generation across 5 manipulation types, with paired (clean, forged) sampling
2️⃣	ViT Detector	DINOv3 ViT-L/16 + LoRA adaptation + EoMT mask head — unified detection & localization
3️⃣	Meta-Harness	Evolving MLLM harness that converts detector outputs into structured forensic reports

Repository Layout

.
├── base_trainer.py                     # Training utilities & metrics
├── cfg.py                              # Runtime configuration
├── ds.py                               # Datasets & dataloaders
├── main.py                             # Train / validation / inference entry point
├── model/
│   ├── eomt_sep_query.py               # Main detector (DINOv3 + LoRA + EoMT)
│   ├── lora.py                         # Single-expert LoRA modules
│   ├── mask_classification_loss.py     # Mask2Former-style loss
│   └── scale_block.py                  # ConvTranspose upscaling block
├── meta_harness/
│   ├── test_submission.py              # Generate challenge-format reports
│   ├── precompute_submission_artifacts.py
│   ├── harness.py                      # Report generation base class
│   ├── llm_clients.py                  # OpenAI-compatible LLM client
│   ├── overlay.py                      # Mask visualization helpers
│   ├── report_utils.py                 # Report formatting utilities
│   ├── template_report_boxreasons_coordspanrepair.py
│   └── config.yaml                     # LLM API configuration
└── TDOC/                               # Auxiliary training & generation modules

⚙️ Environment Setup

# Create a Python 3.10 environment, then install dependencies
pip install -r requirements.txt

📊 Data Preparation

Dataset	Description	Link
RealText-V2	Original challenge dataset	vankey/RealText-V2
RealText-V2-Syn25k	Our synthetic data	Jason37437/RealText-V2-Syn25k
Cross-domain test sets	T-SROIE, OSTF, TPIC-13, RTM	Google Drive
Model Checkpoint	SEED (LoRA rank-1, DINOv3 ViT-L)	Jason37437/SEED / Google Drive

🚀 Training

# 1. Edit cfg.py → mode='train'
# 2. Set your data paths and GPU count
python main.py

📈 Evaluation

# 1. Edit cfg.py → mode='val', eval_mode=['loc','det']
python main.py

📝 Report Generation

The meta_harness/ pipeline converts detector outputs → structured Markdown forensic reports.

# 🔑 Set your OpenAI-compatible API key
export LINKAPI_API_KEY="your-api-key-here"

# 🖼️  Step 1: Precompute overlays, bounding boxes, data URIs
python meta_harness/precompute_submission_artifacts.py

# 📝 Step 2: Generate reports via MLLM
python meta_harness/test_submission.py