python tools/merge_poe_lora.py \
    --base-model checkpoints/qwen3-4b-base-sft-qwen3-8b \
    --adapter checkpoints/qwen3-4b-poe-distill-lora-opd-ppo-clip-locking-b-5-self-distill/checkpoint-100 \
    --output-dir checkpoints/qwen3-4b-poe-distill-lora-opd-ppo-clip-locking-b-5-self-distill-100 \
    --dtype bfloat16