| ---
|
| library_name: transformers
|
| license: other
|
| license_name: lfm1.0
|
| license_link: LICENSE
|
| language:
|
| - en
|
| - ar
|
| - zh
|
| - fr
|
| - de
|
| - ja
|
| - ko
|
| - es
|
| pipeline_tag: text-generation
|
| tags:
|
| - liquid
|
| - unsloth
|
| - lfm2.5
|
| - edge
|
| - mermaid
|
| - diagram-generation
|
| - text-to-diagram
|
| - lfm
|
| base_model:
|
| - unsloth/LFM2.5-1.2B-Instruct
|
| datasets:
|
| - ali-thowfeek/text-to-mermaid
|
| ---
|
|
|
| # Charty-1B
|
|
|
| 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](https://huggingface.co/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](https://huggingface.co/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](https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct) by Liquid AI and is released under the **[LFM Open License v1.0](https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct/blob/main/LICENSE)**.
|
|
|
| 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](../../blob/main/LICENSE) with any redistribution.
|
| - You must retain all copyright and attribution notices.
|
|
|
| See the full [LICENSE](../../blob/main/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](https://www.liquid.ai/) — LFM2.5 base model
|
| - [Unsloth](https://github.com/unslothai/unsloth) — 2× faster fine-tuning framework
|
| - [Celiadraw](https://huggingface.co/Celiadraw) — Original text-to-mermaid dataset source
|
| - [Hugging Face](https://huggingface.co/) — TRL library and model hosting |
