---
license: cc-by-nc-4.0
language:
- en
base_model:
- Qwen/Qwen2.5-VL-7B-Instruct
pipeline_tag: image-text-to-text
library_name: transformers
tags:
- chart-to-code
- multimodal
- vision-language
- sft
- cold-start
- matplotlib
---

# MM-ReCoder-SFT-Cold-Start

<p align="center">
  <a href="https://cvpr.thecvf.com/Conferences/2026"><b>CVPR 2026</b></a>
  &nbsp;|&nbsp;
  <a href="https://zitiantang.github.io/MM-ReCoder/">Project Page</a>
  &nbsp;|&nbsp;
  <a href="https://arxiv.org/abs/2604.01600">arXiv</a>
  &nbsp;|&nbsp;
  <a href="https://github.com/ZitianTang/MM-ReCoder">Code</a>
  &nbsp;|&nbsp;
  <a href="https://huggingface.co/cwbc/MM-ReCoder">Final RL Model</a>
</p>

**MM-ReCoder-SFT-Cold-Start** is the supervised fine-tuned cold-start
checkpoint released alongside the CVPR 2026 paper
[*MM-ReCoder: Advancing Chart-to-Code Generation with Reinforcement Learning and Self-Correction*](https://arxiv.org/abs/2604.01600).
It is fine-tuned from
[`Qwen/Qwen2.5-VL-7B-Instruct`](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct)
to bootstrap the chart-to-code and self-correction behaviors before the
multi-turn RL stages.

> **This is an intermediate checkpoint**, not the final MM-ReCoder model.
> If you want the best chart-to-code performance, use
> [`cwbc/MM-ReCoder`](https://huggingface.co/cwbc/MM-ReCoder) instead.
> This checkpoint is released for researchers who want to reproduce or
> ablate the RL stages of the paper.

## Intended Use

This checkpoint is intended as the **starting point for multi-turn RL**
training. The pipeline is:

1. **SFT cold-start** *(this checkpoint)* — Qwen2.5-VL-7B-Instruct fine-tuned
   on chart-to-code demonstrations.
2. **Multi-turn RL (GRPO), stage 1** — shared-first-turn optimization,
   initialized from this checkpoint.
3. **Multi-turn RL (GRPO), stage 2** — full-trajectory optimization, resumed
   from stage 1. The result is released as
   [`cwbc/MM-ReCoder`](https://huggingface.co/cwbc/MM-ReCoder).

## Usage

To kick off RL from this cold-start checkpoint, clone the
[official repository](https://github.com/ZitianTang/MM-ReCoder) and run the
stage 1 training script (which references this checkpoint via
`REF_MODEL_PATH=cwbc/MM-ReCoder-SFT-Cold-Start`):

```bash
git clone https://github.com/ZitianTang/MM-ReCoder.git
cd MM-ReCoder
# Follow the Installation section in the repo README, then launch the
# LLM-as-a-judge reward server (see the RL Training section).

# Stage 1: multi-turn GRPO with a shared first turn.
bash examples/mmrecoder/train/stage1-shared-first-turn.sh

# Stage 2: multi-turn GRPO on the full trajectory, resumed from stage 1.
bash examples/mmrecoder/train/stage2-full-trajectory.sh
```

### Multi-Turn Inference with the Cold-Start Model

This checkpoint also supports the multi-turn self-correction inference
loop from the repository — useful for measuring the RL gains over the
SFT-only baseline. Reuse the inference scripts and override the model path:

```bash
# Download the cold-start checkpoint.
hf download cwbc/MM-ReCoder-SFT-Cold-Start

# Two-turn self-correction on ChartMimic, using the cold-start model.
bash examples/mmrecoder/inference/chartmimic_2turns.sh \
    model.path=cwbc/MM-ReCoder-SFT-Cold-Start \
    data.output_path=generations/coldstart_chartmimic_2turns.json
```

The self-correction *policy* is sharpened by the RL stages, so the
cold-start model will generally underperform [`cwbc/MM-ReCoder`](https://huggingface.co/cwbc/MM-ReCoder)
on multi-turn benchmarks; this is the intended baseline comparison.

### Direct single-turn use

You can also load the checkpoint directly with `transformers` to inspect
single-turn chart-to-code behavior:

```python
from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration
import torch

model_id = "cwbc/MM-ReCoder-SFT-Cold-Start"
processor = AutoProcessor.from_pretrained(model_id)
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    model_id, torch_dtype=torch.bfloat16, device_map="auto"
)
```

## Citation

```bibtex
@inproceedings{tang2026mmrecoder,
    title={MM-ReCoder: Advancing Chart-to-Code Generation with Reinforcement Learning and Self-Correction},
    author={Zitian Tang and Xu Zhang and Jianbo Yuan and Yang Zou and Varad Gunjal and Songyao Jiang and Davide Modolo},
    booktitle={CVPR},
    year={2026}
}
```

## License

Released under the Apache 2.0 License, inheriting from the base
Qwen2.5-VL-7B-Instruct license.