README.md

---
base_model: Qwen/Qwen2.5-VL-3B-Instruct
library_name: peft
pipeline_tag: image-text-to-text
tags:
- vision
- vqa
- qwen2.5-vl
- lora
- transformers
license: apache-2.0
---

# VQA Base Model

Fine-tuned VQA model using Qwen2.5-VL-3B-Instruct with LoRA.

**Performance:**
- **Validation Accuracy: 88.69%** (345/389)
- **High-res (512px) Accuracy: 89.72%** (349/389)
- Baseline model for the project

**Part of 3-Model Ensemble:**
- Combined with Improved Epoch 1 and Improved Epoch 2
- **Ensemble Validation: 90.75%**
- **Ensemble Test (Kaggle): 91.82%**

## Model Details

- **Base Model:** Qwen/Qwen2.5-VL-3B-Instruct
- **Fine-tuning Method:** LoRA (Low-Rank Adaptation)
- **Quantization:** 4-bit (NF4)
- **Hardware:** NVIDIA A100 40GB
- **Training:** Fine-tuned on VQA dataset (604 samples)

## LoRA Configuration

```python
{
    "r": 16,
    "lora_alpha": 32,
    "lora_dropout": 0.05,
    "target_modules": [
        "q_proj", "k_proj", "v_proj", "o_proj",
        "gate_proj", "up_proj", "down_proj"
    ]
}
```

## Usage

```python
from transformers import AutoModelForVision2Seq, AutoProcessor, BitsAndBytesConfig
from peft import PeftModel
import torch

# Load model with 4-bit quantization
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16
)

base_model = AutoModelForVision2Seq.from_pretrained(
    "Qwen/Qwen2.5-VL-3B-Instruct",
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True
)

model = PeftModel.from_pretrained(base_model, "ikellllllll/vqa-base-model")
processor = AutoProcessor.from_pretrained(
    "Qwen/Qwen2.5-VL-3B-Instruct",
    min_pixels=512*512,
    max_pixels=512*512,
    trust_remote_code=True
)

# IMPORTANT: Set left-padding for decoder-only models
processor.tokenizer.padding_side = 'left'
```

## Inference Settings

- **Image Resolution:** 512×512px (higher resolution recommended)
- **Batch Size:** 32 (for A100 40GB)
- **Padding:** Left-padding (critical for decoder-only models!)

## Dataset

- **Training:** 604 VQA samples
- **Validation:** 389 VQA samples
- **Test:** 3,887 VQA samples

## Performance Notes

- 384px resolution: 88.69% validation accuracy
- 512px resolution: 89.72% validation accuracy (+1.03%)
- **Higher resolution significantly improves performance**

## Links

- **GitHub Repository:** [SSAFY_AI_competition](https://github.com/ikellllllll/SSAFY_AI_competition)
- **Related Models:**
  - [vqa-improved-epoch1](https://huggingface.co/ikellllllll/vqa-improved-epoch1) (90.49%)
  - [vqa-improved-epoch2](https://huggingface.co/ikellllllll/vqa-improved-epoch2) (90.23%)

## Citation

```bibtex
@misc{vqa-base-model,
  author = {Team 203},
  title = {VQA Base Model},
  year = {2025},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/ikellllllll/vqa-base-model}}
}
```

## License

Apache 2.0