mumu-0011/NeFo

Evaluating and Enhancing Negation Comprehension in Remote Sensing MLLMs

Haochen Han, Jue Wang, Alex Jinpeng Wang, Fangming Liu
_{Peng Cheng Laboratory, Tsinghua University, Central South University}

Overview

This repository hosts the NeFo Qwen2.5-VL LoRA adapter outputs for evaluating and enhancing negation comprehension in remote sensing multimodal large language models (MLLMs). NeFo is built on top of LLaMA-Factory and focuses on negation-aware visual question answering and related remote sensing vision-language tasks.

The uploaded adapter is for the VQA sample-150 setting and is intended to be used with the Qwen2.5-VL-7B-Instruct base model.

Resources

• Project code: https://github.com/mumu011/NeFo
• Benchmark dataset: https://huggingface.co/datasets/mumu-0011/RS-Neg
• Base model: https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct

Model Files

The adapter and related run artifacts are stored in:

nefo_qwen2_5vl_sample_150/

Key files include:

• adapter_model.safetensors and adapter_config.json: LoRA adapter weights and configuration.
• tokenizer.json, tokenizer_config.json, preprocessor_config.json, and video_preprocessor_config.json: tokenizer and processor files used for the run.
• predict-temperature_0.0-max_new_tokens_512/generated_predictions_rank0.jsonl: generated predictions from the evaluation run.
• checkpoint-1/ and checkpoint-4/: saved training checkpoints and trainer artifacts.
• logfile.txt and efficiency_stats.txt: run logs and efficiency statistics.

This repository does not contain the full Qwen2.5-VL base model weights. Please load the base model separately and apply the LoRA adapter from this repository.

Quick Start

Install the required packages in your NeFo/LLaMA-Factory environment, then load the adapter with PEFT:

from peft import PeftModel
from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration

base_model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    "Qwen/Qwen2.5-VL-7B-Instruct",
    torch_dtype="auto",
    device_map="auto",
)

model = PeftModel.from_pretrained(
    base_model,
    "mumu-0011/NeFo",
    subfolder="nefo_qwen2_5vl_sample_150",
)

processor = AutoProcessor.from_pretrained(
    "mumu-0011/NeFo",
    subfolder="nefo_qwen2_5vl_sample_150",
)

For full training, inference, and evaluation scripts, please refer to the project repository.

Citation

Thanks to the open-source code of LLaMA-Factory.

If you find this work useful, please cite the related paper:

@article{han2026evaluating,
  title={Evaluating and Enhancing Negation Comprehension in Remote Sensing MLLMs},
  author={Han, Haochen and Wang, Jue and Wang, Alex Jinpeng and Liu, Fangming},
  journal={arXiv preprint arXiv:2606.20177},
  year={2026}
}