README.md

---
license: cc-by-nc-4.0
language:
- en
base_model:
- deepseek-ai/deepseek-coder-6.7b-instruct
- google/siglip-so400m-patch14-384
pipeline_tag: image-text-to-text
---

# Aligned Multi-View Scripts for Universal Chart-to-Code Generation

CharLuMA-1.3B generates a plotting script in Python, R, or LaTeX from a chart image. This 1.3B variant is from the paper *"Aligned Multi-View Scripts for Universal Chart-to-Code Generation"* (ACL 2026 Main Conference).

- **Paper:** https://arxiv.org/abs/2604.24559
- **Code (required for inference):** https://github.com/zhihan72/CharLuMA
- **6.7B Variant:** [CharLuMA-6.7B](https://huggingface.co/Zhihan/CharLuMA-6.7B)
- **Training data:** [Chart2NCode](https://huggingface.co/datasets/Zhihan/Chart2NCode)

| Backbone | Vision encoder | Output languages | dtype |
|---|---|---|---|
| DeepSeek-Coder-1.3B-Instruct | SigLIP-SO400M-patch14-384 | Python, R, LaTeX | bfloat16 |

## Usage

This is a custom architecture, so `AutoModel.from_pretrained` will not load it. 

Clone the codebase and use its loader:

```bash
git clone https://github.com/Zhihan72/CharLuMA
```

Before loading, replace the placeholder paths in `config.json` (`/your_local_path/...`) with `deepseek-ai/deepseek-coder-1.3b-instruct` (or a local copy) and `google/siglip-so400m-patch14-384` (or a local copy). 

Inference example: `scripts/inference_charluma.py` in the repo.

## Citation

If you find our work useful, consider citing our paper as follows:

```bibtex
@misc{zhang2026aligned,
  title         = {Aligned Multi-View Scripts for Universal Chart-to-Code Generation},
  author        = {Zhihan Zhang and Lizi Liao},
  year          = {2026},
  eprint        = {2604.24559},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2604.24559}
}
```