junwatu/candi-sailor2-8b-gguf

Candi Sailor2 8B GGUF

candi-sailor2-8b-gguf is a GGUF release of Candi Sailor2 8B, a fine-tuned Sailor2 model for Indonesian cultural heritage, temples (candi), and Javanese history.

This repository is intended for local inference with tools such as llama.cpp, Ollama, LM Studio, Jan, and other GGUF-compatible runtimes.

Model Details

• Base model: sail/Sailor2-8B
• Merged model: junwatu/candi-sailor2-8b
• LoRA adapter: junwatu/candi-sailor2-8b-lora
• Fine-tuning method: LoRA SFT, then merged into the base model
• Training quantization: 4-bit NF4
• GGUF quantizations: Q4_K_M and Q5_K_M
• Main languages: Indonesian and Javanese, with inherited multilingual support from Sailor2

Available Files

File	Quantization	Suggested use
candi-sailor2-Q4_K_M.gguf	Q4_K_M	Smaller file, good default for most local use
candi-sailor2-Q5_K_M.gguf	Q5_K_M	Larger file, better quality if you have enough RAM or VRAM

Quick Start With Ollama

Create a Modelfile:

FROM ./candi-sailor2-Q4_K_M.gguf

SYSTEM "You are Candi, a helpful assistant specialized in Indonesian cultural heritage, temples (candi), and Javanese history. Answer in clear Indonesian unless the user asks for another language."

PARAMETER temperature 0.7
PARAMETER top_p 0.9

Then run:

ollama create candi-sailor2 -f Modelfile
ollama run candi-sailor2

Example prompt:

Apa itu Candi Borobudur?

Quick Start With llama.cpp

Download one GGUF file, then run:

llama-cli -m candi-sailor2-Q4_K_M.gguf \
  -p "Apa itu Candi Prambanan?"

To serve an OpenAI-compatible local API:

llama-server -m candi-sailor2-Q4_K_M.gguf

Intended Use

This model is designed for:

• Indonesian cultural heritage chatbots
• Candi and Javanese history explanation
• Indonesian-language educational assistants
• Local AI demos using GGUF runtimes
• RAG-based assistants that retrieve exact candi facts from a trusted dataset

Important Limitation

For exact facts such as coordinates, address, elevation, site condition, and geo-validation status, use this model with RAG or another trusted source.

The fine-tune improves behavior and answer style, but it should not be treated as a database. If the answer needs exact factual accuracy, retrieve the source record first and ask the model to explain it.

Training Notes

The LoRA training setup used:

Parameter	Value
LoRA rank	16
LoRA alpha	32
LoRA dropout	0.05
Learning rate	2e-4
Epochs	2
Max sequence length	2048
Optimizer	paged_adamw_8bit
Scheduler	cosine
GPU	L40S

The behavior goal was to make the assistant more careful with uncertain geographic records and more consistent in Indonesian cultural-heritage answers.

Example Prompts

Jelaskan Candi Borobudur dengan bahasa sederhana.

Apa perbedaan candi Hindu dan candi Buddha?

Kalau data lokasi candi punya geo_flag reverse_geo_needs_review, apakah aman ditulis valid tanpa catatan?

Ceritakan legenda Roro Jonggrang secara singkat.

License

Apache 2.0, following the base Sailor2 model license.

Citation

@misc{junwatu2026candi_sailor2_gguf,
  title={Candi Sailor2 8B GGUF},
  author={junwatu},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/junwatu/candi-sailor2-8b-gguf}
}

@article{sailor2,
  title={Sailor2: Advancing Multilingual Large Language Models for Southeast Asian Languages},
  author={Sea AI Lab},
  journal={arXiv preprint arXiv:2502.12982},
  year={2025}
}