disentangled-vla/libero-island-ablation-v12-sigreg-medium
V12 — Single head + SIGReg medium weight (λ=0.01) Bridges V3 (λ_sig=0.1, sig-dominated) and V8 (λ_sig=0.00025, action-dominated). Architecture / Training Backbone: Qwen2.5-VL-3B-Instruct + LoRA (r=32, q/v_proj) Single AttentiveLatentHead A + ResNet action head Loss: 1.0·L1_action + 0.01·SIGReg(z_a, z_domain_current) 100K steps, bs=64, lr=1e-4 cosine, 10 epochs Results — epoch_10 Setting SR AA 0.72 BB 0.72 C_A 0.12 C_B 0.24 AL_B… See the full description on the dataset page: https://huggingface.co/datasets/disentangled-vla/libero-island-ablation-v12-sigreg-medium.
0315
V12 — Single head + SIGReg medium weight (λ=0.01)
Bridges V3 (λsig=0.1, sig-dominated) and V8 (λsig=0.00025, action-dominated).
Architecture / Training
- Backbone: Qwen2.5-VL-3B-Instruct + LoRA (r=32, q/v_proj)
- Single AttentiveLatentHead A + ResNet action head
- Loss:
1.0·L1_action + 0.01·SIGReg(z_a, z_domain_current) - 100K steps, bs=64, lr=1e-4 cosine, 10 epochs
Results — epoch_10
Videos: 25 ep × 6 settings = 150 mp4.