📄 S1-Parser: Efficient Multimodal Document Parsing via Dynamic Structure-Semantics Fusion

🔗 Codebase  

S1-Parser is a highly efficient multimodal text parsing tool designed to enable accurate and efficient parsing of complex documents. Instead of relying solely on static fine-tuning or single-stage optimization, it employs a strategy of first Supervised Fine-Tuning (SFT) then Reinforcement Learning (RL), effectively fine-tuning the model on critical aspects such as formula syntax correctness, symbol integrity, and structural rationality—balancing parsing precision and efficiency across diverse document types.

---

📰 News

• [2025/10/28] We release the Code for S1-Parser.

🚀 Features

• 🧩 Supervised Fine-Tuning with task-oriented ([Parse Target: Scientific Equations]) to sharpen domain adaptation.
• 🎯 Multi-stage RL to refine, stabilize, and accelerate the learning process in strategic of behaviors.
• 📊 Benchmarked on Scientific Literature Dataset: SCI_LLM

---

🛢️ Benchmarks

• sci_llm
• LaTeX_OCR

---

⚙️ Environment Setup (Recommended)

We recommend using Python 3.10 and PyTorch ≥ 2.7.

Install the environment:

# Recommend Python 3.10.18
git clone https://github.com/ScienceOne-AI/S1-Parser.git
cd S1-Parser
pip install -r requirements.txt

🏋️ Training

S1-Parser training proceeds in two stages with different designs:

# Stage 1: Execute the Supervised Fine-Tuning (SFT) to acquire fundamental LaTeX OCR.
bash scripts/run_train_ocr_sft_model.sh

# Stage 2: Execute the GRPO training to optimize LaTeX formula syntax, symbol and structure.
bash scripts/run_train_ocr_grpo_model.sh

---

Make sure to configure your model paths and data in script/run_train_ocr_*.sh.

---

🔍 Acknowledgements

We build and reference on the following open source trunks, and thank the following sources for their contributions to the open source community:

• Qwen2.5-VL
• olmOCR
• MonkeyOCR
• MistralOCR
• Dolphin