ddlx-track3-final-assets

Model Weights

This repository tracks code, configs, and documentation. Large model weights should be downloaded from the GitHub Release assets, Hugging Face, or the organizer verification bundle.

Required Weights

Classifier

• Expected path: models/classifier/convnextb_cls_dev_adapt_head_stage4_last.pt
• Model: ConvNeXt-B face-level binary classifier
• Image-level aggregation: max(face_fake_probability)
• Final fake threshold: 0.20

Detectors

• Expected path: models/detectors/old_fullmask_continue96_stageb3_best.pt
• Expected path: models/detectors/repeat2_conservative_lr1e4_stageb_best.pt
• Expected path: models/detectors/yolov8m512_stageab_stageb_best.pt

These three detector streams are fused with WBF using configs/final_wbf.yaml.

Explanation Model

• Base VLM expected path: models/explanation/qwen2_5_vl_3b_instruct/
• LoRA adapter expected path: models/explanation/qwen2_5_vl_3b_lora_checkpoint1500/
• Adapter checkpoint: checkpoint-1500
• Key adapter file: adapter_model.safetensors

The final selected WBF submission reused cached Qwen-generated text artifacts for exact zip reconstruction. Method-level verification should rerun the Qwen base model plus the LoRA adapter on a small subset or full metadata.

Recommended Release Layout

v1.0-final/
  convnextb_cls_dev_adapt_head_stage4_last.pt
  old_fullmask_continue96_stageb3_best.pt
  repeat2_conservative_lr1e4_stageb_best.pt
  yolov8m512_stageab_stageb_best.pt
  qwen2_5_vl_3b_lora_checkpoint1500.tar.gz
  text_sources.tar.gz
  detector_pred_boxes.json
  DDLX_Track3_FinalVerification_WBF_20260530_1431.tar

If the Qwen2.5-VL-3B-Instruct base model is not attached to the Release, download it from the official model source and place it at the expected path above.