metadata
license: apache-2.0
datasets:
- lmms-lab/LLaVA-OneVision-Data
language:
- en
- zh
metrics:
- accuracy
library_name: transformers
tags:
- multimodal
model-index:
- name: llava-onevision-qwen-72b-si
results:
- task:
type: multimodal
dataset:
type: ai2d
name: AI2D
metrics:
- name: accuracy
type: accuracy
value: 85.1
verified: true
- task:
type: multimodal
dataset:
type: chartqa
name: ChartQA
metrics:
- name: accuracy
type: accuracy
value: 84.9
verified: true
- task:
type: multimodal
dataset:
type: docvqa
name: DocVQA
metrics:
- name: accuracy
type: accuracy
value: 93.5
verified: true
- task:
type: multimodal
dataset:
type: infovqa
name: InfoVQA
metrics:
- name: accuracy
type: accuracy
value: 77.7
verified: true
- task:
type: multimodal
dataset:
type: mathverse
name: MathVerse
metrics:
- name: accuracy
type: accuracy
value: 37.7
verified: true
- task:
type: multimodal
dataset:
type: mathvista
name: MathVista
metrics:
- name: accuracy
type: accuracy
value: 66.5
verified: true
- task:
type: multimodal
dataset:
type: mmbench
name: MMBench
metrics:
- name: accuracy
type: accuracy
value: 86.6
verified: true
- task:
type: multimodal
dataset:
type: mme
name: MME
metrics:
- name: score
type: score
value: 2269
verified: true
- task:
type: multimodal
dataset:
type: mmmu
name: MMMU
metrics:
- name: accuracy
type: accuracy
value: 57.4
verified: true
- task:
type: multimodal
dataset:
type: mmvet
name: MMVet
metrics:
- name: accuracy
type: accuracy
value: 60
verified: true
- task:
type: multimodal
dataset:
type: mmstar
name: MMStar
metrics:
- name: accuracy
type: accuracy
value: 65.2
verified: true
- task:
type: multimodal
dataset:
type: seed-bench
name: Seed-Bench
metrics:
- name: accuracy
type: accuracy
value: 77.6
verified: true
- task:
type: multimodal
dataset:
type: science-qa
name: Science-QA
metrics:
- name: accuracy
type: accuracy
value: 91.3
verified: true
- task:
type: multimodal
dataset:
type: imagedc
name: ImageDC
metrics:
- name: accuracy
type: accuracy
value: 91.5
verified: true
- task:
type: multimodal
dataset:
type: mmlbench
name: MMLBench
metrics:
- name: accuracy
type: accuracy
value: 84.4
verified: true
- task:
type: multimodal
dataset:
type: realworldqa
name: RealWorldQA
metrics:
- name: accuracy
type: accuracy
value: 73.8
verified: true
- task:
type: multimodal
dataset:
type: vibe-eval
name: Vibe-Eval
metrics:
- name: accuracy
type: accuracy
value: 46.7
verified: true
- task:
type: multimodal
dataset:
type: llava-w
name: LLaVA-W
metrics:
- name: accuracy
type: accuracy
value: 93.7
verified: true
- task:
type: multimodal
dataset:
type: l-wilder
name: LLaVA-Wilder
metrics:
- name: accuracy
type: accuracy
value: 72.9
verified: true
pipeline_tag: image-text-to-text
LLaVA-OneVision
Play with the model on the LLaVA OneVision Chat.
Table of Contents
Model Summary
The LLaVA-OneVision models are 0.5/7/72B parameter models trained on LLaVA-OneVision, based on Qwen2 language model with a context window of 32K tokens.
- Repository: LLaVA-VL/LLaVA-NeXT
- Project Website: llava-onevision.lmms-lab.com
- Paper: LLaVA-OneVision
- Point of Contact: Bo Li
- Languages: English, Chinese
Use
Intended use
The model was trained on LLaVA-OneVision Dataset and have the ability to interact with images, multi-image and videos.
Feel free to share your generations in the Community tab!
Generation
# pip install git+https://github.com/LLaVA-VL/LLaVA-NeXT.git
from llava.model.builder import load_pretrained_model
from llava.mm_utils import get_model_name_from_path, process_images, tokenizer_image_token
from llava.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN, IGNORE_INDEX
from llava.conversation import conv_templates, SeparatorStyle
from PIL import Image
import requests
import copy
import torch
import sys
import warnings
warnings.filterwarnings("ignore")
pretrained = "lmms-lab/llava-onevision-qwen2-0.5b-si"
model_name = "llava_qwen"
device = "cuda"
device_map = "auto"
tokenizer, model, image_processor, max_length = load_pretrained_model(pretrained, None, model_name, device_map=device_map) # Add any other thing you want to pass in llava_model_args
model.eval()
url = "https://github.com/haotian-liu/LLaVA/blob/1a91fc274d7c35a9b50b3cb29c4247ae5837ce39/images/llava_v1_5_radar.jpg?raw=true"
image = Image.open(requests.get(url, stream=True).raw)
image_tensor = process_images([image], image_processor, model.config)
image_tensor = [_image.to(dtype=torch.float16, device=device) for _image in image_tensor]
conv_template = "qwen_1_5" # Make sure you use correct chat template for different models
question = DEFAULT_IMAGE_TOKEN + "\nWhat is shown in this image?"
conv = copy.deepcopy(conv_templates[conv_template])
conv.append_message(conv.roles[0], question)
conv.append_message(conv.roles[1], None)
prompt_question = conv.get_prompt()
input_ids = tokenizer_image_token(prompt_question, tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt").unsqueeze(0).to(device)
image_sizes = [image.size]
cont = model.generate(
input_ids,
images=image_tensor,
image_sizes=image_sizes,
do_sample=False,
temperature=0,
max_new_tokens=4096,
)
text_outputs = tokenizer.batch_decode(cont, skip_special_tokens=True)
print(text_outputs)
Training
Model
- Architecture: SO400M + Qwen2
- Pretraining Stage: LCS-558K, 1 epoch, projector
- Mid Stage: A mixture of 4.7M high-quality synthetic data, 1 epoch, full model
- Final-Image Stage: A mixture of 3.6M single-image data, 1 epoch, full model
- OneVision Stage: A mixture of 1.6M single-image/multi-image/video data, 1 epoch, full model
- Precision: bfloat16
Hardware & Software
- GPUs: 256 * Nvidia Tesla A100 (for whole model series training)
- Orchestration: Huggingface Trainer
- Neural networks: PyTorch
Citation
@article{li2024llavaonevision,
title={LLaVA-OneVision},
}