Robo4D-v5-Qwen2.5-VL-7B-LoRA

Robo4D v5 Qwen2.5-VL-7B LoRA

This is a PEFT LoRA adapter for Qwen/Qwen2.5-VL-7B-Instruct, fine-tuned on the Robo4D v5 robot-manipulation video QA task. It is not a standalone checkpoint.

Intended use

The adapter is intended for four-choice video questions about a robot-held object's direction, distance, or speed during manipulation. Direction labels may be relative to the robot base or the camera, depending on the prompt.

Load it with the exact base model and a PEFT-compatible Qwen2.5-VL inference path. The training answer format is a single choice letter.

Training

• Training set: 18,000 episode-capped Robo4D v5 examples.
• Sampling: balanced by question type, source view, and answer class where data availability permits.
• One epoch; LoRA rank 16, alpha 32, dropout 0.05.
• Base model: Qwen/Qwen2.5-VL-7B-Instruct.

On the held-out Robo4D v5 eval set (2,300 questions), this adapter scored 56.52% overall. This is an in-domain, episode-disjoint result.

Limitations

Robo4D training clips are deliberately filtered for grounded, sustained held-object motion. This adapter should not be assumed to improve broad video reasoning or premise checking, including cases with no visible target or no meaningful motion. Early VLM4D transfer checks showed degradation relative to the base model; treat it as a task-specific research adapter.

Loading

from transformers import Qwen2_5_VLForConditionalGeneration
from peft import PeftModel

base = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    "Qwen/Qwen2.5-VL-7B-Instruct", device_map="auto"
)
model = PeftModel.from_pretrained(base, "cpsu04/Robo4D-v5-Qwen2.5-VL-7B-LoRA")

The adapter contains no training videos or question data. Use of the underlying model and any downstream dataset remains subject to their respective licenses and terms.