Liyama-3B

🦙 Liyama-3B

Liyama-3B is a fine-tuned version of Meta’s LLaMA-3B (3.2) model, built to understand and respond fluently in Tagalog. It was trained on the AnoNa dataset over 3 epochs, aiming for natural, context-aware instruction-following in Filipino.

---

🔤 Origin of the Name

The name Liyama is a Tagalified version of llama, reflecting both its LLaMA base and its Tagalog-focused language capabilities. It mirrors how Filipino often adapts foreign terms into familiar, phonetic forms—like camera → kamera, lion → leon, and now, llama → liyama.

---

🧠 Training Data: The AnoNa Dataset

Liyama-3B was trained solely on response completions from the AnoNa dataset — a self-instruct corpus generated using Gemini 1.5 and 2.0.

Inspired by SimpleQnA, the dataset contains short, helpful instruction-response pairs. But AnoNa introduces several improvements:

• ✅ Less English, More Tagalog prompts
• ✅ Less IFEVAL-style formatting
• ✅ No overuse of modifiers in instructions
• ✅ Balanced task types to avoid dominant categories
• ✅ Complex tasks favored (65% complex / 35% simple)
• ✅ Reduced sycophancy and generic praise
• ✅ Improved follow-up handling
• ✅ AI self-intro appears only when relevant
• ✅ Implicit chain-of-thought reasoning, not labeled
• ✅ Extra task types added to increase variety

This focus creates a model that's practical, straightforward, and tuned for realistic conversational use in Filipino, without excessive formatting or irrelevant disclaimers.

---

🗣️ Use Case

Liyama-3B is ideal for:

• Answering questions in Tagalog
• Writing essays, reflections, and letters in Filipino
• Following natural instructions, even when mixed with English
• Chat-based tasks where fluency and tone matter
• Educational or community apps centered around local language use

---

📦 Model Details

---

Liyama-3B is part of a broader effort to create open, practical Filipino-language models for real use—not just benchmarks. Expect follow-ups tuned for multi-turn chat, reasoning, and creative tasks.