bmh201708/qwen2vl-7b-lora-categories

Qwen2-VL-7B Category LoRA Adapters

This repository contains LoRA (Low-Rank Adaptation) adapters for the Qwen2-VL-7B-Instruct model, fine-tuned on different task categories.

Model Description

These LoRA adapters were trained using federated learning approaches on category-specific datasets from the FedMABench benchmark. Each category contains:

• v0: Initial training checkpoint (基础训练后的初始LoRA)
• global_lora_10: Global aggregated LoRA after 10 rounds of federated learning (第10轮联邦学习聚合后的全局LoRA)

Available Categories

Category	v0 Checkpoint	Global LoRA (Round 10)	Training Version
Entertainment	✅ checkpoint-387	✅ global_lora_10	v0-20260131-213955
Lives	✅ checkpoint-500	✅ global_lora_10	v0-20260201-080741
Office	✅ checkpoint-1000	✅ global_lora_10	v0-20260131-211441
Shopping	✅ checkpoint-1000	✅ global_lora_10	v0-20260201-062221
Traveling	✅ checkpoint-500	✅ global_lora_10	v0-20260131-211143

Directory Structure

qwen2vl-7b-lora-categories/
├── Entertainment/
│   ├── v0/                 # Initial checkpoint LoRA
│   └── global_lora_10/     # Round 10 federated LoRA
├── Lives/
│   ├── v0/
│   └── global_lora_10/
├── Office/
│   ├── v0/
│   └── global_lora_10/
├── Shopping/
│   ├── v0/
│   └── global_lora_10/
└── Traveling/
    ├── v0/
    └── global_lora_10/

Usage

Loading with PEFT

from peft import PeftModel
from transformers import Qwen2VLForConditionalGeneration

# Load base model
base_model = Qwen2VLForConditionalGeneration.from_pretrained(
    "Qwen/Qwen2-VL-7B-Instruct",
    torch_dtype="auto",
    device_map="auto"
)

# Load a specific category LoRA (e.g., Entertainment global_lora_10)
model = PeftModel.from_pretrained(
    base_model,
    "bmh201708/qwen2vl-7b-lora-categories",
    subfolder="Entertainment/global_lora_10"
)

Loading with Hugging Face Hub

from huggingface_hub import snapshot_download

# Download specific category LoRA
local_path = snapshot_download(
    repo_id="bmh201708/qwen2vl-7b-lora-categories",
    allow_patterns=["Entertainment/global_lora_10/*"]
)

Base Model

• Model: Qwen/Qwen2-VL-7B-Instruct
• LoRA Library: PEFT 0.17.1

Training Details

These LoRAs were trained as part of the FedMABench (Federated Mobile Agent Benchmark) project for mobile GUI agent tasks. Each category represents different application domains:

• Entertainment: Games, media, entertainment apps
• Lives: Daily life, utility apps
• Office: Productivity, document, work-related apps
• Shopping: E-commerce, shopping apps
• Traveling: Travel, navigation, tourism apps

License

Please refer to the Qwen2-VL license for usage terms.

Citation

If you use these LoRA adapters, please cite the FedMABench project.