pdufour/Llama-3.2-11B-Vision-Instruct-WebSight

Model Card for Llama-3.2-11B-Vision-WebSight

LLama 3.2 Vision Instruct trained on 10k samples from https://huggingface.co./datasets/HuggingFaceM4/WebSight.

Model Details

Model Description

• Developed by: pdufour
• Model type: Vision Language Model
• Language(s) (NLP): English
• License: MIT
• Finetuned from model: meta-llama/Llama-3.2-11B-Vision-Instruct

How to Get Started with the Model

from transformers import AutoModelForVision2Seq, AutoTokenizer, AutoProcessor
from peft import PeftModel
from PIL import Image
import torch

model = PeftModel.from_pretrained(
    AutoModelForVision2Seq.from_pretrained("meta-llama/Llama-3.2-11B-Vision-Instruct", device_map="auto", load_in_4bit=True),
    "pdufour/Llama-3.2-11B-Vision-WebSight"
)
tokenizer = AutoTokenizer.from_pretrained("pdufour/Llama-3.2-11B-Vision-WebSight")
processor = AutoProcessor.from_pretrained("meta-llama/Llama-3.2-11B-Vision-Instruct")

inputs = processor(text="Generate code for a web page that looks exactly like this. <|image|>", images=Image.open("fashion.jpg"), return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(input_ids=inputs['input_ids'], max_new_tokens=4096, do_sample=True, temperature=0.7, top_p=0.9)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training Details

Training Data

Vision-language dataset used for instruction tuning.

Training Procedure

Training Hyperparameters

• Training regime: Fine-tuning with LoRA
• Learning rate: 0.0002
• Batch size: 10
• Gradient accumulation steps: 1
• Number of epochs: 3.0
• Optimizer: adamw_torch_fused
• LR scheduler type: constant
• Weight decay: 0.0
• FP16 Training: False

Speeds, Sizes, Times

• Training Duration: Unknown hours
• Number of Parameters: Unknown trainable parameters
• Model Size: 0.08 GB

Evaluation

Metrics

Results

• epoch: 0.9000
• grad_norm: 0.2568
• learning_rate: 0.0002
• loss: 0.0791
• step: 900.0000

Technical Specifications

Model Architecture and Objective

LoRA-tuned Vision-Language Model based on Llama architecture.

Compute Infrastructure

• Hardware Type: GPU
• Number of GPUs: 1

Software

• Framework versions:
  • PEFT 0.13.2
  • PyTorch 2.5.0+cu121

Model Card Contact

For questions about this model, please file an issue on the GitHub repository.