qwen3_vl_8b_rmbench_taco_lora_b16_10k

High-level planner for RMBench (9 simulated tabletop tasks): Qwen/Qwen3-VL-8B-Instruct fine-tuned with LoRA on the per-tick planning records of Myungkyu/RMBench-taco-gemini — demonstrations with dense subtask labels from the task-specific context (offline annotation).

• LoRA adapter (PEFT) for Qwen/Qwen3-VL-8B-Instruct: r=32, alpha=64, dropout=0.0 on the language-model q/k/v/o/gate/up/down projections; step 10000 of a 10000-step run (optimizer state not included)
• Contract: recent head-camera frames + task goal + carried memory text → JSON with the current subtask, the updated memory, a keyframe flag with caption and a retrieval query
• Optimizer batch 16, 10000 steps, learning rate 2e-4; this repo holds the adapter of step 10000 only (load with peft.PeftModel.from_pretrained(base, repo) on top of Qwen/Qwen3-VL-8B-Instruct, or merge for inference)