---
title: HalDet-llava-13b
canonical_url: "https://www.modelscope.cn/models/ZJUNLP/HalDet-llava-13b"
md_url: "https://www.modelscope.cn/models/ZJUNLP/HalDet-llava-13b.md"
repository: ZJUNLP/HalDet-llava-13b
last_updated: 2026-06-15
license: mit
parameters: 13.4B
tensor_type:
  - F16
library_name:
  - safetensors
downloads: 52
stars: 0
---

# HalDet-llava-13b

> HalDet-llava-13b - ZJUNLP 在 ModelScope 开源的模型。多模态幻觉检测模型

ZJUNLP/HalDet-llava-13b 是 ModelScope 魔搭社区上的 13.4B 参数机器学习模型，采用 mit 许可。

- **Repository**: ZJUNLP/HalDet-llava-13b
- **License**: mit
- **Parameters**: 13.4B
- **Downloads**: 52
- **Stars**: 0
- **Last updated**: 2026-06-15

Source: https://www.modelscope.cn/models/ZJUNLP/HalDet-llava-13b

---

## HalDet-LLaVA

HalDet-LLaVA is designed for multimodal hallucination detection, trained on the MHaluBench training dataset, achieving detection performance close to that of using GPT4-Vision.

HalDet-LLaVA is trained on the [MHaluBench training set](https://huggingface.co/datasets/openkg/MHaluBench/blob/main/MHaluBench_train.json) using LLaVA-v1.5, specific parameters can be found in the file [finetune_task_lora.sh](https://github.com/zjunlp/EasyDetect/blob/main/HalDet-LLaVA/finetune_task_lora.sh).

We trained HalDet-LLaVA on 1-A800 in 1 hour. If you don"t have enough GPU resources, we will soon provide model distributed training scripts.

You can inference our HalDet-LLaVA by using [inference.py](https://github.com/zjunlp/EasyDetect/blob/main/HalDet-LLaVA/inference.py)

To view more detailed information about HalDet-LLaVA and the train dataset, please refer to the [EasyDetect](https://github.com/zjunlp/EasyDetect) and [readme](https://github.com/zjunlp/EasyDetect/blob/main/HalDet-LLaVA/README.md)