---
license: apache-2.0
base_model:
- Qwen/Qwen3-VL-4B-Thinking
pipeline_tag: image-text-to-text
library_name: transformers
metrics:
- accuracy
tags:
- vision-language
- chart-question-answering
- multimodal
- pytorch
model-index:
- name: Supertron-VL-4B
  results:
  - task:
      type: image-text-to-text
      name: Chart Question Answering
    dataset:
      name: ChartQA
      type: HuggingFaceM4/ChartQA
      split: test
    metrics:
    - name: ChartQA relaxed accuracy
      type: accuracy
      value: 0.7891
    - name: Exact match
      type: accuracy
      value: 0.7109
---

# **Supertron-VL-4B: A Chart-Focused Vision-Language Model**

## **Model Description**

**Supertron-VL-4B** is a vision-language model fine-tuned from **Qwen/Qwen3-VL-4B-Thinking** for chart understanding and chart question answering. It reads chart images, extracts values, compares visual elements, and answers concise questions about plotted data.

* **Developed by:** Surpem
* **Model type:** Vision-Language Model
* **Architecture:** Qwen3-VL dense multimodal transformer, 4B class
* **Fine-tuned from:** [Qwen/Qwen3-VL-4B-Thinking](https://huggingface.co/Qwen/Qwen3-VL-4B-Thinking)
* **License:** Apache 2.0

---

## **Evaluation**

Local Modal H100 benchmark using the Hugging Face `transformers` `image-text-to-text` pipeline:

| Benchmark | Split | Samples | Exact Accuracy | Relaxed ChartQA Accuracy |
|---|---:|---:|---:|---:|
| ChartQA | test | 256 | 0.7109 | 0.7891 |

**Note:** This is an offline local benchmark, not an official Hugging Face leaderboard verification.

---

## **Get Started**

```python
from transformers import AutoProcessor, AutoModelForImageTextToText
from PIL import Image
import torch

model_id = "Surpem/Supertron-VL-4B"

processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)

image = Image.open("chart.png").convert("RGB")
question = "What is the highest value shown in the chart?"
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": image},
            {
                "type": "text",
                "text": (
                    "Read the chart image and answer the question concisely. "
                    "Return only the final answer, without chain-of-thought.\n"
                    f"Question: {question}"
                ),
            },
        ],
    }
]

text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(text=[text], images=[image], padding=True, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=48, do_sample=False)
generated = outputs[:, inputs["input_ids"].shape[1]:]
print(processor.batch_decode(generated, skip_special_tokens=True)[0].strip())
```

---

## **Limitations**

Supertron-VL-4B is specialized for chart question answering. It may make mistakes on crowded charts, ambiguous labels, color-only questions, arithmetic-heavy questions, or charts with very small text.