Model Details

• Model family: Gemma 4 (E2B)
• Base model: unsloth/gemma-4-E2B-it
• Primary language: Tetun
• Fine-tuning: LoRA (rank 64), merged into the base weights
• Training data: 138,033 Tetun conversations with 261,593 question–answer exchanges
• Quantization: Q4_K_M (3.4 GB)
• Context length: 4,096 tokens by default
• Input: Text only

Usage

Requires a recent version of Ollama with Gemma 4 support. The model needs about 4–5 GB of free memory and also runs on CPU-only laptops.

Run the model with Ollama:

ollama run labadainofficial/labadain-e2b-tetun:q4_k_m --think=false

Keep thinking off with --think=false. The training conversations contain no reasoning traces, so the model was not tuned to think before answering.

API

curl http://localhost:11434/api/chat -d '{
  "model": "labadainofficial/labadain-e2b-tetun:q4_k_m",
  "messages": [{"role": "user", "content": "Saida mak kapitál Timor-Leste nian?"}],
  "think": false,
  "stream": false
}'

Python

pip install ollama

from ollama import chat

response = chat(
    model="labadainofficial/labadain-e2b-tetun:q4_k_m",
    messages=[
        {"role": "user", "content": "To'o iha-ne'ebé ona progresu edukasaun iha Timor-Leste?"},
    ],
    think=False,
)

print(response.message.content)

Parameters

The low default temperature gives focused, consistent answers. For more varied writing, raise it to 0.7–1.0:

• In ollama run: /set parameter temperature 1.0
• In the API: "options": {"temperature": 1.0}
• In Python: chat(..., options={"temperature": 1.0})

Tips

• A system prompt is optional. The model was fine-tuned without one. A short instruction such as “Responde iha Tetun” can still help keep answers in Tetun.
• Multi-turn chat works. Over half of the training conversations have two or more question–answer exchanges.

Training Data

The model was fine-tuned on 138,033 Tetun conversations (roughly 102 million tokens) compiled by Labadain, including some material machine-translated into Tetun. Together they contain 261,593 question–answer exchanges:

Multi-turn conversations have up to 33 exchanges. Duplicates were removed before training, and 3,000 conversations were held out for validation.

Training

The model was fine-tuned with Unsloth and TRL’s SFTTrainer for up to 2 epochs, keeping the checkpoint with the lowest validation loss. Loss was computed only on the model’s answers. The merged model was then converted to GGUF and quantized to Q4_K_M. Full training details are on the Hugging Face model card.

Intended Uses

• Tetun conversational AI
• Tetun text generation
• Research on multilingual and low-resource NLP

Limitations

• Factual errors. This is a small model and can state incorrect facts, dates, names and figures with confidence, including about Timor-Leste. Check important information against reliable sources.
• Fixed knowledge. The model only knows what was in the base model and the training data. It has no access to current events.
• Long inputs. Fine-tuning covered conversations of up to 4,096 tokens. Quality on longer conversations or documents has not been tested.
• Tool calling. Ollama lists tool support because it comes with Gemma 4, but the training data contains no tool calls, so tool use has not been tested.
• Quantization. This 4-bit version can be slightly less accurate than the original bfloat16 model.
• No evaluation yet. The model has not been evaluated on standard benchmarks, downstream applications, or for safety beyond what the base model provides.

Hugging Face

Merged model: https://huggingface.co/labadain/labadain-E2B-it-Tetun

GGUF: https://huggingface.co/labadain/labadain-E2B-it-Tetun-GGUF

License

This model is released under the Apache 2.0 license, following the Gemma 4 license of the base model.

About Labadain

Labadain is a research initiative advancing information retrieval (IR), natural language processing (NLP), and artificial intelligence (AI) for Tetun and other low-resource language settings.

Related Projects

• Labadain Chat — Conversational AI for the Tetun Language
• Labadain Search - A Search Engine for the Tetun Language