gui360-v17-dual-aux-sft-warmup

V17 Dual-Aux SFT Warmup

Cooperative LoRA (r=128, T=2 comm rounds) trained with dual-aux format SFT warmup.

Purpose

Teaches the model the dual-aux output format before RL training:

<aux_a>[Expert A perspective]</aux_a>
<aux_b>[Expert B perspective]</aux_b>
<decision><tool_call>{action}</tool_call></decision>

Training Details

• Base model: Qwen2.5-VL-7B-Instruct
• Method: Cooperative LoRA from scratch, 1 epoch SFT on template-based dual-aux data
• Data: GUI-360 balanced 2K episodes (16,837 train steps), template-based aux content
• Architecture: r=128, T=2 communication rounds, target modules: q/k/v/o_proj
• Training: 2 nodes x 4 GPUs, grad_accum=8, lora_lr=2e-5, route_lr=1e-3
• Result: Train loss 1.01 -> 0.035, val loss 0.049

Role in V17 Pipeline

1. This checkpoint (SFT warmup) — teaches format structure
2. V17 RL (next) — phase-conditional hard routing learns aux content
   • aux_a -> Expert A (route=0.9), aux_b -> Expert B (route=0.1), decision -> learned routing
   • Diversity loss encourages expert heterogeneity

Files

• — LoRA A1, A2, B weights (~504M)
• — Communication layer weights (~3.6M)
• — Routing weights (~0.2M)
• — Architecture config
• — Training state (step, epoch)

Citation

Part of the Cooperative LoRA research for GUI agents.