Kendamarron/Llama-3.2-11B-Vision-Instruct-Swallow-8B-Merge-v0.1

Model Information

Kendamarron/Llama-3.2-11B-Vision-Instruct-Swallow-8B-Mergeの初期バージョンです。

Llama-3.1シリーズの代わりにLlama-3シリーズを使用しています。

Llama-3.1を使用したモデルと体感の出力はあまり変わりません。

Detail

https://zenn.dev/kendama/articles/280a4089cb8a72

Recipe

Llama-3.2-11B-Vision-Instruct + (Llama-3-Swallow-8B-v0.1 - Meta-Llama-3-8B)

• Vision Model: meta-llama/Llama-3.2-11B-Vision-Instruct
• Base Text Model: meta-llama/Meta-Llama-3-8B
• Japanese Text Model: tokyotech-llm/Llama-3-Swallow-8B-v0.1

License

Llama 3.2 Community License

How to use

import requests
import torch
from PIL import Image
from transformers import MllamaForConditionalGeneration, AutoProcessor

model_id = "Kendamarron/Llama-3.2-11B-Vision-Instruct-Swallow-8B-Merge-v0.1"

model = MllamaForConditionalGeneration.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
processor = AutoProcessor.from_pretrained(model_id)

url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/0052a70beed5bf71b92610a43a52df6d286cd5f3/diffusers/rabbit.jpg"
image = Image.open(requests.get(url, stream=True).raw)

messages = [
    {"role": "user", "content": [
        {"type": "image"},
        {"type": "text", "text": "この画像で一句詠んでください。"}
    ]}
]
input_text = processor.apply_chat_template(messages, add_generation_prompt=True)
inputs = processor(
    image,
    input_text,
    add_special_tokens=False,
    return_tensors="pt"
).to(model.device)

output = model.generate(**inputs, max_new_tokens=30)
print(processor.decode(output[0]))