---
license: mit
base_model: Qwen/Qwen3-VL-4B-Instruct
library_name: peft
pipeline_tag: image-text-to-text
tags:
- lora
- peft
- qwen3-vl
- robotics
- subgoal-prediction
---

# hl-planner-simple-subgoal

LoRA adapter on **Qwen3-VL-4B-Instruct** for robot subgoal prediction from
setup video + task goal.

- Base: `Qwen/Qwen3-VL-4B-Instruct`
- Adapter: LoRA (~132 MB safetensors)
- License: MIT

## Inference env vars (required)

These MUST match training; setting them higher (or leaving the defaults)
silently degrades quality.

```bash
export IMAGE_MAX_TOKEN_NUM=128
export VIDEO_MAX_TOKEN_NUM=128
export FPS_MAX_FRAMES=10
```

## Use

Load via `peft` on top of Qwen3-VL with the env vars above set BEFORE
importing `swift.llm` / `transformers`.