---
license: apache-2.0
language:
- en
- zh
pipeline_tag: visual-question-answering
tags:
- multimodal
library_name: transformers
---

<div align=center><img src ="https://cdn-uploads.huggingface.co/production/uploads/6299c90ef1f2a097fcaa1293/XEfp5nnJOixkGAOyF8UtN.png"/></div>  


## Introduction

**Xinyuan-VL-2B** is a high-performance multimodal large model for the end-side from the Cylingo Group, which is fine-tuned with `Qwen/Qwen2-VL-2B-Instruct`, and uses more than 5M of multimodal data as well as a small amount of plain text data. 

It performs well on several authoritative Benchmarks.

## How to use

In order to rely on the thriving ecology of the open source community, we chose to fine-tune [Qwen/Qwen2-VL-2B-Instruct](https://huggingface.co/Qwen/Qwen2-VL-2B-Instruct) to form our `Cylingo/Xinyuan-VL- 2B`. 

Thus, using `Cylingo/Xinyuan-VL-2B` is consistent with using `Qwen/Qwen2-VL-2B-Instruct`:

```Python
from transformers import Qwen2VLForConditionalGeneration, AutoTokenizer, AutoProcessor
from qwen_vl_utils import process_vision_info

# default: Load the model on the available device(s)
model = Qwen2VLForConditionalGeneration.from_pretrained(
    "Cylingo/Xinyuan-VL-2B", torch_dtype="auto", device_map="auto"
)

# default processer
processor = AutoProcessor.from_pretrained("Cylingo/Xinyuan-VL-2B")

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg",
            },
            {"type": "text", "text": "Describe this image."},
        ],
    }
]

# Preparation for inference
text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
)
inputs = inputs.to("cuda")

# Inference: Generation of the output
generated_ids = model.generate(**inputs, max_new_tokens=128)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print(output_text)

```



## Evaluation
We evaluated **[XinYuan-VL-2B](https://huggingface.co/thomas-yanxin/XinYuan-VL-2B)** using the [VLMEvalKit](https://github.com/open-compass/VLMEvalKit) toolkit across the following benchmarks and found that **XinYuan-VL-2B** **outperformed** [Qwen/Qwen2-VL-2B-Instruct](https://huggingface.co/Qwen/Qwen2-VL-2B-Instruct) released by Alibaba Cloud, as well as other models of comparable parameter scale that have significant influence in the open-source community.

<p align="center">
  <img src="https://cdn-uploads.huggingface.co/production/uploads/6299c90ef1f2a097fcaa1293/7ThTCYfd_lDzsvaFLlUv2.png">
</p>

You can see the results in [opencompass/open_vlm_leaderboard](https://huggingface.co/spaces/opencompass/open_vlm_leaderboard):

| Benchamrk | MiniCPM-2B | InternVL-2B | Qwen2-VL-2B | **XinYuan-VL-2B** |
| :---: | :---: | :---: | :---: | :---: |
| MMB-CN-V11-Test | 64.5 | 68.9 | 71.2 | **74.3** |
| MMB-EN-V11-Test | 65.8 | 70.2 | 73.2 | **76.5** |
| MMB-EN | 69.1 | 74.4 | 74.3 | **78.9** |
| MMB-CN | 66.5 | 71.2 | 73.8 | **76.12** |
| CCBench | 45.3 | 74.7 | 53.7 | 55.5 |
| MMT-Bench | 53.5 | 50.8 | 54.5 | **55.2** |
| RealWorld | 55.8 | 57.3 | 62.9 | **63.9** |
| SEEDBench\_IMG | 67.1 | 70.9 | 72.86 | **73.4** |
| AI2D | 56.3 | 74.1 | **74.7** | 74.2 |
| MMMU | 38.2 | 36.3 | **41.1** | 40.9 |
| HallusionBench | 36.2 | 36.2 | 42.4 | **55.00** |
| POPE | 86.3 | 86.3 | 86.82 | **89.42** |
| MME | 1808.6 | **1876.8** | 1872.0 | 1854.9 |
| MMStar | 39.1 | 49.8 | 47.5 | **51.87** |
| SEEDBench2\_Plus | 51.9 | 59.9 | 62.23 | **62.98** |
| BLINK | 41.2 | 42.8 | **43.92** | 42.98 |
| OCRBench | 605 | 781 | **794** | 782 |
| TextVQA | 74.1 | 73.4 | **79.7** | 77.6 |