VincentNi/wan22-rollout-put-bottles-dustbin-step0
收藏资源简介:
Wan2.2 TI2V Step-0 Rollout — RoboTwin `put_bottles_dustbin`数据集包含160个视频(10个场景×16个样本),这些视频是由Wan2.2-TI2V-5B模型和合并的Vidar LoRA在RoboTwin `put_bottles_dustbin`任务(依次将3个瓶子放入垃圾桶)上生成的。这些视频是预NFT训练(step-0)的基线样本,用于评估奖励模型的行为和种子RL微调。生成配置包括基础模型、LoRA、采样器、步数、位移、CFG比例、帧数和种子等详细信息。数据集布局为每个场景包含16个视频文件,总计160个MP4文件(约602 MB)。
The Wan2.2 TI2V Step-0 Rollout — RoboTwin `put_bottles_dustbin` dataset consists of 160 video rollouts (10 scenes × 16 samples) generated by the Wan2.2-TI2V-5B model and merged Vidar LoRA on the RoboTwin `put_bottles_dustbin` task (putting 3 bottles into a dustbin sequentially). These are the pre-NFT-training (step-0) baseline samples used to evaluate reward-model behavior and seed RL fine-tuning. The generation config includes details such as the base model, LoRA, sampler, steps, shift, CFG scale, frames, and seeds. The dataset layout includes 16 video files per scene, totaling 160 MP4 files (~602 MB).




