paligemma2-3b-lora-vqa-v21-enhanced-d8000-r8 - v2.1 Enhanced

This is a v2.1 Enhanced LoRA adapter for PaliGemma-2 3B trained on VQA tasks.

🆕 v2.1 Enhanced Improvements

• EOS Token Learning: Explicit EOS tokens for better generation termination
• Memory Optimization: 16-step gradient accumulation for stability
• VizWiz Format Support: Full support with most frequent answer selection
• Robust Label Masking: Enhanced prompt masking during training
• Production Memory Management: Advanced garbage collection

Usage

from transformers import AutoProcessor, PaliGemmaForConditionalGeneration
from peft import PeftModel
import torch
from PIL import Image

# Base model
base_model_id = "google/paligemma2-3b-mix-224"
adapter_id = "yu3733/paligemma2-3b-lora-vqa-v21-enhanced-d8000-r8"

# Load processor
processor = AutoProcessor.from_pretrained(base_model_id)

# Load base model with quantization (optional)
model = PaliGemmaForConditionalGeneration.from_pretrained(
    base_model_id,
    torch_dtype=torch.float16,
    device_map="auto"
)

# Load LoRA adapter
model = PeftModel.from_pretrained(model, adapter_id)

# Prepare input
image = Image.open("your_image.jpg")
prompt = "<image>\nQuestion: What is in this image?\nAnswer:"

# Process
inputs = processor(text=prompt, images=image, return_tensors="pt")
inputs = {k: v.to(model.device) for k, v in inputs.items()}

# Generate
with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=20)

# Decode
print(processor.decode(outputs[0], skip_special_tokens=True))

Training Configuration

• Base Model: google/paligemma2-3b-mix-224
• LoRA Rank: 8
• Training Framework: PEFT + Transformers
• Optimization: 4-bit quantization + gradient checkpointing
• Dataset: VizWiz VQA

License

Same as the base model (see google/paligemma2-3b-mix-224)