SynhalaAI/SynhalaAI-Llama-3.2-3B-CPT-Merged
4B • Updated • 61
SynhalaAI
non-profit
AI & ML interests
Sri Lanka's first dedicated AI initiative for the Sinhala language — building open models, datasets, benchmarks, and tools for 22 million Sinhala speakers. Powered by skAI — SKY PRODUCTION.
Recent Activity
View all activity
Organization Card
🦁 SynhalaAI
Sinhala. Intelligence. Future.
Powered by skAI — SKY PRODUCTION
What is SynhalaAI?
SynhalaAI is Sri Lanka's first dedicated Artificial Intelligence initiative built exclusively for the Sinhala language.
This is not a single model. This is not a one-time experiment.
SynhalaAI is a long-term commitment to building the complete AI infrastructure that Sinhala has always deserved — models, datasets, benchmarks, tools, and open resources for every developer, researcher, and creator who wants to build with us.
"AI should work for everyone — in every language. SynhalaAI is our commitment to making that true for Sinhala." — SKY PRODUCTION
Why Sinhala?
22 million people speak Sinhala — yet the language has almost no dedicated support in modern AI technology.
Most of the world's leading AI systems are built for high-resource languages: English, Mandarin, Spanish. Languages like Sinhala receive little to no dedicated investment, creating a growing digital divide where entire communities are left behind as the world accelerates forward with AI.
SynhalaAI exists to close that gap.
What We're Building
| Project | Status | Description |
|---|---|---|
| SynhalaAI-Tharushi-3B | 🔜 Coming Soon | Fine-tuned LLaMA 3.2 3B for Sinhala — runs offline on mobile |
| SynhalaEes | 🔄 In Progress | Sri Lanka's first dedicated Sinhala AI benchmark |
| Sinhala Dataset | 🔄 In Progress | 10K–100K instruction/response pairs for Sinhala NLP |
| Sinhala GGUF Models | 🔜 Coming Soon | Quantized models (q4_k_m, q5_k_m, q8_0) for mobile & edge devices |
SynhalaAI-Tharushi-3B — First Model
Our first release under SynhalaAI is a fine-tuned LLaMA 3.2 3B model optimized for the Sinhala language using Unsloth + QLoRA.
Key Facts
- 🇱🇰 First offline Sinhala AI model that runs on a mobile phone — no internet required
- 📱 Mobile-friendly — GGUF format for llama.cpp / Ollama
- 🤗 Open source — free for everyone under the SynhalaAI Open License
- 📊 Evaluated on SynhalaEes — Sri Lanka's first Sinhala AI benchmark
Capabilities
- ✅ Sinhala chat & instruction following
- ✅ Sinhala text generation
- ✅ Sinhala Q&A
- ✅ English ↔ Sinhala translation
- ✅ Fully offline — no cloud, no internet required
SynhalaEes Benchmark
SynhalaEes is Sri Lanka's first dedicated Sinhala language AI benchmark to measure and improve AI performance on the Sinhala language.
Because you can't improve what you don't measure.
Benchmark results and methodology will be published alongside the first model release.
License
All models and datasets released under the SynhalaAI Open License v1.0
✅ Free to use — personal, research, and commercial
✅ Free to modify and build upon
✅ Attribution required — cite "SynhalaAI by SKY PRODUCTION"
❌ Cannot harm or misrepresent the Sinhala language or culture
❌ Cannot be rebranded as a new base model without disclosure
❌ SKY PRODUCTION branding cannot be used without permission
Roadmap
- Project launch — SynhalaAI brand & initiative
- SynhalaAI-Tharushi-3B model release (safetensors + GGUF)
- SynhalaEes benchmark publication
- Sinhala instruction dataset release
- SynhalaAI API for developers
- SynhalaAI v2 — expanded capabilities
- Community dataset contributions open
Connect With Us
| Platform | Link |
|---|---|
| 🐙 GitHub | github.com/SynhalaAI |
SynhalaAI — Sinhala. Intelligence. Future.
Powered by skAI — SKY PRODUCTION 🦁🇱🇰