Mungert/Logics-Parsing-GGUF

<d

Logics-Parsing GGUF Models

Model Generation Details

This model was generated using llama.cpp at commit c8dedc99.

---

Quantization Beyond the IMatrix

I've been experimenting with a new quantization approach that selectively elevates the precision of key layers beyond what the default IMatrix configuration provides.

In my testing, standard IMatrix quantization underperforms at lower bit depths, especially with Mixture of Experts (MoE) models. To address this, I'm using the --tensor-type option in llama.cpp to manually "bump" important layers to higher precision. You can see the implementation here:
👉 Layer bumping with llama.cpp

While this does increase model file size, it significantly improves precision for a given quantization level.

I'd love your feedback—have you tried this? How does it perform for you?

---

Click here to get info on choosing the right GGUF model format

---

iv align="center">

🤗 GitHub   |   🤖 Demo   |   📑 Technical Report

Introduction

report	chemistry	paper	handwritten

Logics-Parsing is a powerful, end-to-end document parsing model built upon a general Vision-Language Model (VLM) through Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL). It excels at accurately analyzing and structuring highly complex documents.

Key Features

• Effortless End-to-End Processing

  • Our single-model architecture eliminates the need for complex, multi-stage pipelines. Deployment and inference are straightforward, going directly from a document image to structured output.
  • It demonstrates exceptional performance on documents with challenging layouts.

• Advanced Content Recognition

  • It accurately recognizes and structures difficult content, including intricate scientific formulas.
  • Chemical structures are intelligently identified and can be represented in the standard SMILES format.

• Rich, Structured HTML Output

  • The model generates a clean HTML representation of the document, preserving its logical structure.
  • Each content block (e.g., paragraph, table, figure, formula) is tagged with its category, bounding box coordinates, and OCR text.
  • It automatically identifies and filters out irrelevant elements like headers and footers, focusing only on the core content.

• State-of-the-Art Performance

  • Logics-Parsing achieves the best performance on our in-house benchmark, which is specifically designed to comprehensively evaluate a model’s parsing capability on complex-layout documents and STEM content.

Benchmark

Existing document-parsing benchmarks often provide limited coverage of complex layouts and STEM content. To address this, we constructed an in-house benchmark comprising 1,078 page-level images across nine major categories and over twenty sub-categories. Our model achieves the best performance on this benchmark.

Model Type	Methods	Overall Edit ↓		Text Edit Edit ↓		Formula Edit ↓		Table TEDS ↑		Table Edit ↓		ReadOrderEdit ↓		ChemistryEdit ↓	HandWritingEdit ↓
		EN	ZH	EN	ZH	EN	ZH	EN	ZH	EN	ZH	EN	ZH	ALL	ALL
Pipeline Tools	doc2x	0.209	0.188	0.128	0.194	0.377	0.321	81.1	85.3	0.148	0.115	0.146	0.122	1.0	0.307
	Textin	0.153	0.158	0.132	0.190	0.185	0.223	76.7	86.3	0.176	0.113	0.118	0.104	1.0	0.344
	mathpix*	0.128	0.146	0.128	0.152	0.06	0.142	86.2	86.6	0.120	0.127	0.204	0.164	0.552	0.263
	PP_StructureV3	0.220	0.226	0.172	0.29	0.272	0.276	66	71.5	0.237	0.193	0.201	0.143	1.0	0.382
	Mineru2	0.212	0.245	0.134	0.195	0.280	0.407	67.5	71.8	0.228	0.203	0.205	0.177	1.0	0.387
	Marker	0.324	0.409	0.188	0.289	0.285	0.383	65.5	50.4	0.593	0.702	0.23	0.262	1.0	0.50
	Pix2text	0.447	0.547	0.485	0.577	0.312	0.465	64.7	63.0	0.566	0.613	0.424	0.534	1.0	0.95
Expert VLMs	Dolphin	0.208	0.256	0.149	0.189	0.334	0.346	72.9	60.1	0.192	0.35	0.160	0.139	0.984	0.433
	dots.ocr	0.186	0.198	0.115	0.169	0.291	0.358	79.5	82.5	0.172	0.141	0.165	0.123	1.0	0.255
	MonkeyOcr	0.193	0.259	0.127	0.236	0.262	0.325	78.4	74.7	0.186	0.294	0.197	0.180	1.0	0.623
	OCRFlux	0.252	0.254	0.134	0.195	0.326	0.405	58.3	70.2	0.358	0.260	0.191	0.156	1.0	0.284
	Gotocr	0.247	0.249	0.181	0.213	0.231	0.318	59.5	74.7	0.38	0.299	0.195	0.164	0.969	0.446
	Olmocr	0.341	0.382	0.125	0.205	0.719	0.766	57.1	56.6	0.327	0.389	0.191	0.169	1.0	0.294
	SmolDocling	0.657	0.895	0.486	0.932	0.859	0.972	18.5	1.5	0.86	0.98	0.413	0.695	1.0	0.927
	Logics-Parsing	0.124	0.145	0.089	0.139	0.106	0.165	76.6	79.5	0.165	0.166	0.136	0.113	0.519	0.252
General VLMs	Qwen2VL-72B	0.298	0.342	0.142	0.244	0.431	0.363	64.2	55.5	0.425	0.581	0.193	0.182	0.792	0.359
	Qwen2.5VL-72B	0.233	0.263	0.162	0.24	0.251	0.257	69.6	67	0.313	0.353	0.205	0.204	0.597	0.349
	Doubao-1.6	0.188	0.248	0.129	0.219	0.273	0.336	74.9	69.7	0.180	0.288	0.171	0.148	0.601	0.317
	GPT-5	0.242	0.373	0.119	0.36	0.398	0.456	67.9	55.8	0.26	0.397	0.191	0.28	0.88	0.46
	Gemini2.5 pro	0.185	0.20	0.115	0.155	0.288	0.326	82.6	80.3	0.154	0.182	0.181	0.136	0.535	0.26

^* Tested on the v3/PDF Conversion API (August 2025 deployment).

Quick Start

1. Installation

conda create -n logis-parsing python=3.10
conda activate logis-parsing

pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cu124

2. Download Model Weights

# Download our model from Modelscope.
pip install modelscope
python download_model.py -t modelscope

# Download our model from huggingface.
pip install huggingface_hub
python download_model.py -t huggingface

3. Inference

python3 inference.py --image_path PATH_TO_INPUT_IMG --output_path PATH_TO_OUTPUT --model_path PATH_TO_MODEL

Acknowledgments

We would like to acknowledge the following open-source projects that provided inspiration and reference for this work:

• Qwen2.5-VL
• OmniDocBench
• Mathpix

---

🚀 If you find these models useful

Help me test my AI-Powered Quantum Network Monitor Assistant with quantum-ready security checks:

👉 Quantum Network Monitor

The full Open Source Code for the Quantum Network Monitor Service available at my github repos ( repos with NetworkMonitor in the name) : Source Code Quantum Network Monitor. You will also find the code I use to quantize the models if you want to do it yourself GGUFModelBuilder

💬 How to test:
Choose an AI assistant type:

• TurboLLM (GPT-4.1-mini)
• HugLLM (Hugginface Open-source models)
• TestLLM (Experimental CPU-only)

What I’m Testing

I’m pushing the limits of small open-source models for AI network monitoring, specifically:

• Function calling against live network services
• How small can a model go while still handling:
  • Automated Nmap security scans
  • Quantum-readiness checks
  • Network Monitoring tasks

🟡 TestLLM – Current experimental model (llama.cpp on 2 CPU threads on huggingface docker space):

• ✅ Zero-configuration setup
• ⏳ 30s load time (slow inference but no API costs) . No token limited as the cost is low.
• 🔧 Help wanted! If you’re into edge-device AI, let’s collaborate!

Other Assistants

🟢 TurboLLM – Uses gpt-4.1-mini :

• **It performs very well but unfortunatly OpenAI charges per token. For this reason tokens usage is limited.
• Create custom cmd processors to run .net code on Quantum Network Monitor Agents
• Real-time network diagnostics and monitoring
• Security Audits
• Penetration testing (Nmap/Metasploit)

🔵 HugLLM – Latest Open-source models:

• 🌐 Runs on Hugging Face Inference API. Performs pretty well using the lastest models hosted on Novita.

💡 Example commands you could test:

1. "Give me info on my websites SSL certificate"
2. "Check if my server is using quantum safe encyption for communication"
3. "Run a comprehensive security audit on my server"
4. '"Create a cmd processor to .. (what ever you want)" Note you need to install a Quantum Network Monitor Agent to run the .net code on. This is a very flexible and powerful feature. Use with caution!

Final Word

I fund the servers used to create these model files, run the Quantum Network Monitor service, and pay for inference from Novita and OpenAI—all out of my own pocket. All the code behind the model creation and the Quantum Network Monitor project is open source. Feel free to use whatever you find helpful.

If you appreciate the work, please consider buying me a coffee ☕. Your support helps cover service costs and allows me to raise token limits for everyone.

I'm also open to job opportunities or sponsorship.

Thank you! 😊