README.md

---
base_model: Qwen/Qwen2.5-VL-3B-Instruct
language:
- fil
- en
license: apache-2.0
tags:
- qwen2.5-vl
- education
- filipino
- fine-tuned
- gguf
---

# KoaliPi SLM — Qwen2.5-VL-3B Fine-tune

KoaliPi SLM is a fine-tuned vision-language model built on top of
[Qwen2.5-VL-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct),
trained for AI-powered STEM study assistance for Filipino students
(high school and college level).

## Files
| File | Size | Description |
|------|------|-------------|
| `koalipi-slm-q4km.gguf` | ~1.9 GB | Q4_K_M quantized — recommended for on-device use |
| `koalipi-slm.gguf` | 3.29 GB | Q8_0 quantized — higher precision |

## Training Details
- **Base model:** Qwen/Qwen2.5-VL-3B-Instruct
- **Method:** LoRA fine-tuning via Unsloth
- **Training steps:** 30
- **Quantization:** 16-bit during training, exported to Q4_K_M via llama.cpp

## Training Data
The model was fine-tuned on a custom KoaliPi dataset covering:
- Handwriting parsing (IAM handwriting dataset + custom Filipino notes)
- Filipino Q&A (Taglish STEM explanations)
- English Q&A (STEM concepts)
- MCQ generation
- Practice problem generation
- Study plan generation

## Instruction Format
No system prompt needed. Uses ChatML format:

**Document parsing:**

Parse this document. Return ONLY a valid JSON with these fields:

topics, difficulty, subject, key_concepts, has_equations, equations, summary.

No explanation, just JSON.

**MCQ generation:**
Generate 3 multiple choice questions about {topic}.

Return ONLY valid JSON array.

**Filipino Q&A:**
{question in Filipino or English}

## Intended Use
- Filipino high school and college STEM students
- On-device inference via llama.rn
- Part of the KoaliPi AI study app

## Limitations
- mmproj (vision encoder) not yet exportable via llama.cpp for Qwen2.5-VL
- Image/VL features require mmproj — text features fully functional
- 30 training steps — suitable for demo, expand dataset for production