llava-llama-3-8b-hf / README.md
Seungyoun's picture
Update "Quick Start"
7e83c59 verified
metadata
license: llama3
library_name: xtuner
datasets:
  - Lin-Chen/ShareGPT4V
pipeline_tag: image-text-to-text

Notice: This repository hosts the xtuner/llava-llama-3-8b-v1_1-hf model, which has been specifically modified to address compatibility issues with the pure transformers library. The original model configuration and index files have been manually adjusted to ensure seamless integration and functionality with the transformers setup. These adjustments have not altered the model weights.


QuickStart

Running with pure transformers library

from transformers import (
    LlavaProcessor,
    LlavaForConditionalGeneration,
)
from PIL import Image
import requests

MODEL_NAME = "Seungyoun/llava-llama-3-8b-hf"

processor = LlavaProcessor.from_pretrained(MODEL_NAME)
processor.tokenizer.add_tokens(
    ["<|image|>", "<pad>"], special_tokens=True
)  # add 128257 <|image|> , <pad>

model = LlavaForConditionalGeneration.from_pretrained(MODEL_NAME).to("cuda:0")
model.resize_token_embeddings(
    len(processor.tokenizer)
)  # resize embeddings for new tokens


# prepare image and text prompt, using the appropriate prompt template
url = "https://upload.wikimedia.org/wikipedia/commons/1/18/Kochendes_wasser02.jpg"
image = Image.open(requests.get(url, stream=True).raw)

template = """<|start_header_id|>system<|end_header_id|>{system_prompt}<|eot_id|>
<|start_header_id|>user<|end_header_id|>{user_msg_1}<|eot_id|>
<|start_header_id|>assistant<|end_header_id|>"""

terminators = [
    processor.tokenizer.eos_token_id,
    processor.tokenizer.convert_tokens_to_ids("<|eot_id|>"),
]

prompt = template.format(
    system_prompt="As a vision-llm, your task is to analyze and describe the contents of the image presented to you. Examine the photograph closely and provide a comprehensive, detailed caption. You should identify and describe the various food items and their arrangement, as well as any discernible textures, colors, and specific features of the containers they are in. Highlight the variety and how these contribute to the overall visual appeal of the meal. Your description should help someone who cannot see the image to visualize its contents accurately.",
    user_msg_1="<|image|>\nGive me detailed description of the image.",
)

inputs = processor(prompt, image, return_tensors="pt").to("cuda:0")

# autoregressively complete prompt
output = model.generate(**inputs, max_new_tokens=1024, eos_token_id=terminators)

print(processor.decode(output[0], skip_special_tokens=False))
# The image captures a moment in a kitchen. The main focus is a white electric kettle, which is plugged in and resting on a black stovetop. The stovetop has four burners, although only one is occupied by the kettle. The background is blurred, drawing attention to the kettle and stovetop. The image does not contain any text or additional objects. The relative position of the objects is such that the kettle is on the stovetop, and the background is blurred.

Model

llava-llama-3-8b-v1_1-hf is a LLaVA model fine-tuned from meta-llama/Meta-Llama-3-8B-Instruct and CLIP-ViT-Large-patch14-336 with ShareGPT4V-PT and InternVL-SFT by XTuner.

Details

Model Visual Encoder Projector Resolution Pretraining Strategy Fine-tuning Strategy Pretrain Dataset Fine-tune Dataset
LLaVA-v1.5-7B CLIP-L MLP 336 Frozen LLM, Frozen ViT Full LLM, Frozen ViT LLaVA-PT (558K) LLaVA-Mix (665K)
LLaVA-Llama-3-8B CLIP-L MLP 336 Frozen LLM, Frozen ViT Full LLM, LoRA ViT LLaVA-PT (558K) LLaVA-Mix (665K)
LLaVA-Llama-3-8B-v1.1 CLIP-L MLP 336 Frozen LLM, Frozen ViT Full LLM, LoRA ViT ShareGPT4V-PT (1246K) InternVL-SFT (1268K)

Results

Image
Model MMBench Test (EN) MMBench Test (CN) CCBench Dev MMMU Val SEED-IMG AI2D Test ScienceQA Test HallusionBench aAcc POPE GQA TextVQA MME MMStar
LLaVA-v1.5-7B 66.5 59.0 27.5 35.3 60.5 54.8 70.4 44.9 85.9 62.0 58.2 1511/348 30.3
LLaVA-Llama-3-8B 68.9 61.6 30.4 36.8 69.8 60.9 73.3 47.3 87.2 63.5 58.0 1506/295 38.2
LLaVA-Llama-3-8B-v1.1 72.3 66.4 31.6 36.8 70.1 70.0 72.9 47.7 86.4 62.6 59.0 1469/349 45.1

Citation

@misc{2023xtuner,
    title={XTuner: A Toolkit for Efficiently Fine-tuning LLM},
    author={XTuner Contributors},
    howpublished = {\url{https://github.com/InternLM/xtuner}},
    year={2023}
}