---
pipeline_tag: text-generation
library_name: transformers
tags:
  - qwen3
  - gguf
  - mandala
  - education
  - fine-tuned
language:
  - ko
  - en
license: apache-2.0
---

# Insighta Mandala v13

Fine-tuned Qwen3-4B model for generating mandala learning plans in JSON format.

## Model Details

- **Base model**: Qwen/Qwen3-4B
- **Fine-tuning**: LoRA on mandala learning plan generation task
- **Languages**: Korean, English
- **Output format**: Structured JSON (mandala chart format)

## Available Quantizations

| Format | Size | Description |
|--------|------|-------------|
| `model.safetensors` | ~8GB | Full F16 weights |
| `insighta-mandala-v13-Q8_0.gguf` | ~4GB | Q8_0 quantized GGUF |
| `insighta-mandala-v13-Q4_K_M.gguf` | ~2.4GB | Q4_K_M quantized GGUF (recommended for CPU) |

## Usage

### With llama-cpp-python

```python
from llama_cpp import Llama

llm = Llama(model_path="insighta-mandala-v13-Q4_K_M.gguf", n_ctx=4096)
output = llm(
    "<|im_start|>user\nTOEFL 100점 만다라트 차트를 만들어줘<|im_end|>\n<|im_start|>assistant\n",
    max_tokens=2048,
    temperature=0.7,
)
print(output["choices"][0]["text"])
```

### With HF Space API

```bash
curl -X POST https://jamesjk4242-insighta-mandala-v13-api.hf.space/api/predict \
  -H "Content-Type: application/json" \
  -d '{"data": ["TOEFL 100점 만다라트 차트를 만들어줘", "You are a helpful assistant that generates mandala learning plans in JSON format.", 2048, 0.7, 0.9, true]}'
```