README.md

---
language:
- en
pipeline_tag: text-generation
---

<p align="center">
    <img src="./Bespoke-Labs-Logo.png" width="550">
</p>

# Bespoke-MiniChart-7B

<a href="https://playground.bespokelabs.ai/minichart">
  <img src="https://cdn-uploads.huggingface.co/production/uploads/6444e4417a7b94ddc2d14e1d/g-QaXrmPLYk5m3Hq5vFtr.png" width="200px" />
</a>

This is an open‑source chart‑understanding Vision‑Language Model (VLM) developed at [Bespoke Labs](https://www.bespokelabs.ai/) and maintained by [Liyan Tang](https://www.tangliyan.com/) and Bespoke Labs. It sets a new state‑of‑the‑art in chart question‑answering (Chart‑QA) for 7 billion‑parameter models, outperforming much larger closed models such as Gemini‑1.5‑Pro and Claude‑3.5 on seven public benchmarks.

1. **Blog Post**: https://www.bespokelabs.ai/blog/bespoke-minichart-7b
2. **Playground**: https://playground.bespokelabs.ai/minichart
---

# Example Outputs

The examples below showcase how Bespoke-MiniChart-7B can perform both visual perception and textual reasoning.


<p align="left">
    <img src="https://cdn-uploads.huggingface.co/production/uploads/6444e4417a7b94ddc2d14e1d/E5WGhi_fVNzCsrKeNeIs3.png" width="700">
</p>

<p align="left">
    <img src="https://cdn-uploads.huggingface.co/production/uploads/6444e4417a7b94ddc2d14e1d/bYKXRm3sfOdX3zd_5qUpK.png" width="700">
</p>


# Model Performance

Bespoke-MiniChart-7B achieves state-of-the-art performance on chart understanding among models with similar sizes. In addition to that, the model can even surpass closed-models such as Gemini-1.5-Pro and Claude-3.5.

<p align="left">
    <img src="https://cdn-uploads.huggingface.co/production/uploads/6444e4417a7b94ddc2d14e1d/5pejAyzPG_tRBU6FwH7PA.png" width="700">
</p>

We also compare the performance of our model finetuned using SFT+DPO vs SFT only. 

In the table below, M1 and M2 are finetuned models with 270K and 1M SFT examples respsectively, and Bespoke-MiniChart-7B is the model finetuned using SFT+DPO. 

<p align="left">
    <img src="https://cdn-uploads.huggingface.co/production/uploads/6444e4417a7b94ddc2d14e1d/WRsPs437niUrXmYtkRajG.png" width="700">
</p>


# Model Use:

[![Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1FEmlwGgn9209iQO-rs2-9UHPLoytwZMH?usp=sharing)

The model is available on the playground here: https://playground.bespokelabs.ai/minichart

You can also run the model with the following snippet: 

```python
import requests
from PIL import Image
from io import BytesIO
import base64
import matplotlib.pyplot as plt
from vllm import LLM, SamplingParams

QA_PROMPT = """Please answer the question using the chart image.

Question: [QUESTION]

Please first generate your reasoning process and then provide the user with the answer. Use the following format:

<think> 
... your thinking process here ... 
</think> 
<answer> 
... your final answer (entity(s) or number) ...
</answer>"""

def get_image_from_url(image_url):
    try:
        response = requests.get(image_url, stream=True)
        response.raise_for_status()
        return Image.open(BytesIO(response.content))
    except Exception as e:
        print(f"Error with image: {e}")
        return None

def get_answer(image_url, question, display=True):
    image = get_image_from_url(image_url)

    if display:
      plt.figure(figsize=(10, 8))
      plt.imshow(image)
      plt.axis('off')
      plt.show()

    if not image:
        return "Error downloading image" 
  
    buffered = BytesIO()
    image.save(buffered, format=image.format or 'JPEG')
    encoded_image = base64.b64encode(buffered.getvalue()).decode('utf-8')
    
    messages = [{
        "role": "user",
        "content": [
            {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{encoded_image}"}},
            {"type": "text", "text": QA_PROMPT.replace("[QUESTION]", question)}
        ]
    }]

    response = llm.chat([messages], sampling_params=SamplingParams(temperature=0, max_tokens=500))
    return response[0].outputs[0].text
    
# Initialize the LLM
llm = LLM(
    model="bespokelabs/Bespoke-MiniChart-7B",
    tokenizer_mode="auto",
    max_model_len=15000,
    tensor_parallel_size=1,
    gpu_memory_utilization=0.9,
    mm_processor_kwargs={"max_pixels": 1600*28*28},
    seed=2025,
    trust_remote_code=True,
)

# Running inference
image_url = "https://github.com/bespokelabsai/minichart-playground-examples/blob/main/images/ilyc9wk4jf8b1.png?raw=true"
question = "How many global regions maintained their startup funding losses below 30% in 2022?"

print("\n\n=================Model Output:===============\n\n", get_answer(image_url, question))
```

---
# Licence

This work is licensed under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/).
For commercial licensing, please contact [email protected].

# Citation

```
@misc{bespoke_minichart_7b,
  title  = {Bespoke-MiniChart-7B: pushing the frontiers of open VLMs for chart understanding},
  author = {Liyan Tang and Shreyas Pimpalgaonkar and Kartik Sharma and Alexandros G. Dimakis and Mahesh Sathiamoorthy and Greg Durrett},
  howpublished = {blog post},
  year   = {2025},
  url={https://huggingface.co/bespokelabs/Bespoke-MiniChart-7B},
}
```

# Acknowledgements

**Bespoke Labs** team:

- Liyan Tang
- Shreyas Pimpalgaonkar
- Kartik Sharma
- Alex Dimakis
- Mahesh Sathiamoorthy
- Greg Durrett


*Model perfected at Bespoke Labs — where careful curation meets cutting‑edge modeling.*