So-Fake-R1 Model Weights

This repository contains the model weights used by So-Fake-R1, an evidence-grounded framework for social-media image forgery detection.

The Qwen2.5-VL-7B-Instruct base model is not redistributed here. The released VLM components are LoRA adapters that should be loaded with Qwen/Qwen2.5-VL-7B-Instruct.

Contents

observer_adapter/          Visual Observer LoRA adapter
decision_maker_adapter/    Decision Maker LoRA adapter
provider/                  Forensic Evidence Provider checkpoints
model_manifest.json        Component manifest
SHA256SUMS                 File checksums

The provider directory contains four small checkpoints:

npr_internalized_feta_mono_blur_only05_e1.pt
feta_v0_L19_cls_only10d_L19_u1_e1.pt
npr_feta_cls_npr_residual_feta10d_e1.pt
l19_heatmap_bbox_head_best.pt

Expected Use

These weights are intended to be used with the So-Fake-R1 inference code. A typical runtime uses:

export QWEN_BASE_MODEL=Qwen/Qwen2.5-VL-7B-Instruct
export SOFAKE_R1_OBSERVER_ADAPTER=/path/to/observer_adapter
export SOFAKE_R1_DM_CKPT=/path/to/decision_maker_adapter
export SOFAKE_R1_PROVIDER_DIR=/path/to/provider

SAM2 is only required when exporting SAM2-refined masks from predicted boxes.

Notes

• Optimizer states, scheduler states, and training logs are intentionally excluded.
• The observer is released as a LoRA adapter rather than a merged full Qwen checkpoint.
• The Decision Maker is the v5.64 route-aware GRPO checkpoint.