KyleZ0906/DreamZero-LIBERO-LoRA

DreamZero LIBERO LoRA

LoRA fine-tuned checkpoint of DreamZero-AgiBot on the LIBERO benchmark.

Base Model

• Backbone: Wan2.1-I2V-14B-480P
• Pretrain: DreamZero-AgiBot (GEAR-Dreams/DreamZero-AgiBot)

Training Details

• Dataset: LIBERO-90 (no-noops), 3921 episodes, 73 tasks
• Method: LoRA (rank=4, alpha=4)
• Trainable params: ~108M / 16.5B total
• Resolution: 320×176 (frame_seqlen=880)
• Action horizon: 24 steps
• Optimizer: AdamW, lr=5e-6, weight_decay=1e-5
• DeepSpeed: ZeRO-2

Usage

Load with DreamZero codebase: https://github.com/GEAR-Dreams/DreamZero