climba/safe-video-cogvideox-refdpo-redirect-lora

Safe-Video CogVideoX Conservative Redirect RefDPO LoRA

This is a CogVideoX-2B LoRA adapter trained with reference-corrected DPO on consensus safety preferences. Preference pairs condition training on the unsafe prompt while using a safer winner video, so this adapter is intended to test unsafe-prompt redirection while preserving video quality.

Intended use

This is the current best-balanced research checkpoint from the local Safe-Video experiments.

Training setup

• Base model: CogVideoX-2B
• Adapter: LoRA rank 64, alpha 64
• Objective: reference-corrected DPO over denoising-loss preference margins
• Steps: 500
• Beta: 10.0
• SFT auxiliary weight: 0.2
• Safety-gap weight: 0.25
• Training prompt mode: unsafe

Local evaluation snapshot

Evaluation used Qwen2-VL as an interim safety and quality judge.

Model	Unsafe rate	Safety cost	Quality	Physics
Base CogVideoX-2B, test40	0.8562	0.6300	0.7601	0.7063
This LoRA, test40	0.8438	0.6216	0.7531	0.7000
Base CogVideoX-2B, benign100	0.3100	0.2430	0.8143	0.7536
This LoRA, benign100	0.3100	0.2392	0.8032	0.7486

Safe-prompt coverage at safety cost <= 0.35 improved from 17/40 to 21/40 while keeping quality close to the base model.

Files

• pytorch_lora_weights.safetensors: Diffusers LoRA adapter weights for the CogVideoX transformer.
• training_summary.json: local training history and final DPO metrics.

Loading

from diffusers import CogVideoXPipeline
import torch

pipe = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-2b", torch_dtype=torch.bfloat16)
pipe.load_lora_weights("climba/safe-video-cogvideox-refdpo-redirect-lora")
pipe.to("cuda")

Caveats

This is a research artifact. Safety scores are judge-model estimates, not a guarantee of safe behavior.