README.md

metadata

language:
  - zh
  - en
license: cc-by-sa-4.0
task_categories:
  - video-text-to-text
tags:
  - multimodal
  - video-understanding
  - short-video
  - benchmark
  - e-commerce
  - vqa
library_name:
  - transformers

[🍎 Home Page] [📖 Technical Report] [\ud83d\udcca Models] [\ud83d\ude80 Demo]

This repository contains KC-MMBench, a new benchmark dataset meticulously tailored for real-world short-video scenarios, as presented in the paper "Kwai Keye-VL Technical Report". Constructed from Kuaishou short video data, KC-MMBench comprises 6 distinct datasets designed to evaluate the performance of Vision-Language Models (VLMs) like Kwai Keye-VL-8B, Qwen2.5-VL, and InternVL in comprehending dynamic, information-dense short-form videos.

For the associated code, detailed documentation, and evaluation scripts, please refer to the official Kwai Keye-VL GitHub repository.

If you want to use KC-MMbench, please download with:

git clone https://huggingface.co/datasets/Kwai-Keye/KC-MMbench

Tasks

Task	Description
CPV	The task of predicting product attributes in e-commerce.
Hot_Videos_Aggregation	The task of determining whether multiple videos belong to the same topic.
Collection_Order	The task of determining the logical order between multiple videos with the same topic.
Pornographic_Comment	The task of whether short video comments contain pornographic content.
High_Like	A binary classification task to determine the rate of likes of a short video.
SPU	The task of determining whether two items are the same product in e-commerce.

Performance

Task	Qwen2.5-VL-3B	Qwen2.5-VL-7B	InternVL-3-8B	MiMo-VL-7B	Kwai Keye-VL-8B
CPV	12.39	20.08	14.95	16.66	55.13
Hot_Videos_Aggregation	42.38	46.35	52.31	49.00	54.30
Collection_Order	36.88	59.83	64.75	78.68	84.43
Pornographic_Comment	56.61	56.08	57.14	68.25	71.96
High_Like	48.85	47.94	47.03	51.14	55.25
SPU	74.09	81.34	75.64	81.86	87.05

Usage

This section provides a quick guide on how to interact with models using the keye-vl-utils library, which is essential for processing and integrating visual language information with Keye Series Models like Kwai Keye-VL-8B.

Install keye-vl-utils

First, install the necessary utility library:

pip install keye-vl-utils

Keye-VL Inference Example

Here's an example of performing inference with a Kwai Keye-VL model, demonstrating how to prepare inputs for both image and video scenarios.

from transformers import AutoModel, AutoProcessor
from keye_vl_utils import process_vision_info

# default: Load the model on the available device(s)
model_path = "Kwai-Keye/Keye-VL-8B-Preview"

model = AutoModel.from_pretrained(
    model_path, torch_dtype="auto", device_map="auto", attn_implementation="flash_attention_2", trust_remote_code=True,
).to('cuda')

# Example messages demonstrating various input types (image, video)
messages = [
    # Image Input Examples
    [{"role": "user", "content": [{"type": "image", "image": "file:///path/to/your/image.jpg"}, {"type": "text", "text": "Describe this image."}]}],
    [{"role": "user", "content": [{"type": "image", "image": "http://path/to/your/image.jpg"}, {"type": "text", "text": "Describe this image."}]}],
    [{"role": "user", "content": [{"type": "image", "image": "data:image;base64,/9j/..."}, {"type": "text", "text": "Describe this image."}]}],
    
    # Video Input Examples (most relevant for KC-MMBench)
    [{"role": "user", "content": [{"type": "video", "video": "file:///path/to/video1.mp4"}, {"type": "text", "text": "Describe this video."}]}],
    [{"role": "user", "content": [{"type": "video", "video": ["file:///path/to/extracted_frame1.jpg", "file:///path/to/extracted_frame2.jpg", "file:///path/to/extracted_frame3.jpg"],}, {"type": "text", "text": "Describe this video."},],}],
    [{"role": "user", "content": [{"type": "video", "video": "file:///path/to/video1.mp4", "fps": 2.0, "resized_height": 280, "resized_width": 280}, {"type": "text", "text": "Describe this video."}]}],
]

processor = AutoProcessor.from_pretrained(model_path)
# Note: model loaded above already
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
images, videos, video_kwargs = process_vision_info(messages, return_video_kwargs=True)
inputs = processor(text=text, images=images, videos=videos, padding=True, return_tensors="pt", **video_kwargs).to("cuda")
generated_ids = model.generate(**inputs)
print(generated_ids)

Evaluation

For detailed instructions on how to evaluate models using the KC-MMBench datasets, including setup and running evaluation scripts, please refer to the evaluation/KC-MMBench/README.md file in the official Kwai Keye-VL GitHub repository.

Below is the example configuration for evaluation using VLMs on our datasets:

{
    "model": "...", # Specify your model
    "data": {
        "CPV": {
            "class": "KwaiVQADataset",
            "dataset": "CPV"
        },
        "Hot_Videos_Aggregation": {
            "class": "KwaiVQADataset",
            "dataset": "Hot_Videos_Aggregation"
        },
        "Collection_Order": {
            "class": "KwaiVQADataset",
            "dataset": "Collection_Order"
        },
        "Pornographic_Comment": {
            "class": "KwaiYORNDataset",
            "dataset": "Pornographic_Comment"
        },
        "High_like":{
            "class":"KwaiYORNDataset",
            "dataset":"High_like"
        },
        "SPU": {
            "class": "KwaiYORNDataset",
            "dataset": "SPU"
        }
    }
}