A compact text-to-Mermaid diagram generation model fine-tuned from Liquid AI's LFM2.5-1.2B-Instruct, built for on-device and mobile deployment. Give it a natural-language description and Charty-1B outputs raw Mermaid syntax only — no explanations, no markdown fences, no conversational filler.

This repository contains the merged full model in 16-bit safetensors format. For quantized GGUF versions (F16 / Q8 / Q4K_M), see ali-thowfeek/Charty-1B-GGUF

🗒️ Model Details

Field	Value
Base model	unsloth/LFM2.5-1.2B-Instruct
Architecture	LFM2 (hybrid: 10 double-gated LIV convolution blocks + 6 GQA attention blocks)
Parameters	1.17B
Context length	32,768 tokens
Vocabulary size	65,536
Fine-tuning method	LoRA (SFT) via Unsloth + Hugging Face TRL
Training dataset	ali-thowfeek/text-to-mermaid (3,837 examples)
Developed by	ali-thowfeek

💬 Chat Template (Unsloth's Fixed Template)

This model ships with Unsloth's fixed Jinja chat template, inherited from the base repo.

The same template is embedded in the GGUF builds (see the GGUF repo) and is applied automatically by llama.cpp with --jinja.

🎯 Intended Use

Generate Mermaid diagram syntax (flowcharts, sequence, class, state, ER, Gantt, pie, git graphs, mindmaps, quadrant charts, …) from plain-text prompts. Run on low-end hardware: mobile phones, edge devices, laptops, embedded systems. Power apps that need offline, private, diagram-as-code generation.

What it does

Input (user prompt): Create a flowchart showing the user login process with MFA verification.

Output (model response):

mermaid

graph TD
    A[User Visits Login Page] --> B[Enter Credentials]
    B --> C{Credentials Valid?}
    C -->|No| D[Show Error Message]
    D --> B
    C -->|Yes| E[Send MFA Code]
    E --> F[Enter MFA Code]
    F --> G{MFA Valid?}
    G -->|No| H[Show MFA Error]
    H --> F
    G -->|Yes| I[Grant Access]

The model outputs only the Mermaid syntax code. No wrapping text, no mermaid fences.

🏃 Inference

Recommended Generation Parameters

Parameter	Value
temperature	0.1
top_k	50
top_p	0.1
repetition_penalty	1.05

📦 GGUF / Quantized Versions

For CPU-only inference, mobile deployment, or use with llama.cpp / Ollama / LM Studio:

👉 ali-thowfeek/Charty-1B-GGUF

Quantization	Use Case
F16	Maximum quality, larger memory footprint
Q8_0	Near-lossless quality, balanced size
Q4KM	Smallest size, best for mobile / edge

📊 Training Details

Detail	Value
Framework	Unsloth + Hugging Face TRL
Method	LoRA Supervised Fine-Tuning (SFT), merged into full weights
Dataset	ali-thowfeek/text-to-mermaid
Dataset size	3,837 examples
Dataset source	Derived from Celiadraw/text-to-mermaid-2, cleaned, reworded, and validated against Mermaid v11 (core)
Validation	100% of training examples produce valid Mermaid v11 core syntax
Output format	Raw Mermaid syntax only (no markdown fences, no explanations)

📜 License

This model is a derivative work of LFM2.5-1.2B-Instruct by Liquid AI and is released under the LFM Open License v1.0.

Key terms:

✅ Free for research, personal, and non-commercial use.
✅ Commercial use permitted for entities with < $10M annual revenue.
❌ Commercial use by entities with ≥ $10M annual revenue is not licensed.
You must include a copy of the LICENSE with any redistribution.
You must retain all copyright and attribution notices.

See the full LICENSE file in this repository for complete terms.

📚 Citation

If you use this model, please cite the base model and this work:

bibtex

@article{liquidai2025lfm2,
  title   = {LFM2 Technical Report},
  author  = {Liquid AI},
  journal = {arXiv preprint arXiv:2511.23404},
  year    = {2025}
}

@misc{thowfeek2026charty,
  title  = {Charty-1B: A Text-to-Mermaid Diagram Generation Model},
  author = {ali-thowfeek},
  year   = {2026},
  url    = {https://huggingface.co/ali-thowfeek/Charty-1B}
}

🙏 Acknowledgements

Liquid AI — LFM2.5 base model
Unsloth — 2× faster fine-tuning framework
Celiadraw — Original text-to-mermaid dataset source
Hugging Face — TRL library and model hosting