climba/safe-video-cogvideox-sft-lora

Safe-Video CogVideoX SFT LoRA

This is a CogVideoX-2B LoRA adapter trained on safety-filtered winner videos from the Commonsense-Safe T2V research workspace.

Intended use

Use this adapter as a quality-preserving supervised fine-tuning baseline for commonsense-safe text-to-video alignment experiments.

Local evaluation snapshot

Evaluation used Qwen2-VL as an interim safety and quality judge over 40 held-out paired prompts with 4 candidates per prompt.

Model	Unsafe rate	Safety cost	Quality	Physics
Base CogVideoX-2B	0.8562	0.6300	0.7601	0.7063
This SFT LoRA	0.8469	0.6272	0.7848	0.7278

This checkpoint is mainly useful as a quality and SFT baseline. It does not strongly reduce unsafe-prompt failures by itself.

Files

• pytorch_lora_weights.safetensors: Diffusers LoRA adapter weights for the CogVideoX transformer.

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-sft-lora")
pipe.to("cuda")

Caveats

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