javanese-qlora

A QLoRA adapter that fine-tunes Gemma 3 4B IT for Javanese (Basa Jawa) conversation and text generation.

• Demo: novanandriyono/javanese (Hugging Face Space, ZeroGPU)
• Developed by: Novan Andriyono (novanandriyono)
• Model type: LoRA adapter (PEFT), not a standalone model
• Language: Javanese (jv)
• Base model: unsloth/gemma-3-4b-it-unsloth-bnb-4bit
• Fine-tuning method: QLoRA (4-bit NF4 base, trained with Unsloth)
• Adapter size: ~600 MB (adapter_model.safetensors)
• License: Gemma Terms of Use. This adapter is a derivative of Gemma and is subject to those terms.

Status: experimental and still in development. Output quality varies, especially for speech levels (ngoko, krama) and rare vocabulary.

Intended use

• Chat and instruction following in Javanese
• Experiments on low-resource language adaptation
• Research and hobby projects

Out-of-scope use

• Anything that needs factual reliability (medical, legal, financial advice)
• Authoritative translation or official documents
• Any use prohibited by the Gemma Prohibited Use Policy

How to use

The adapter was trained on the 4-bit base, but it can be applied to the bf16 base (unsloth/gemma-3-4b-it) and merged for inference. This is what the demo Space does.

Option 1: bf16 base + merge (needs ~10 GB VRAM)

import torch
from transformers import AutoTokenizer, Gemma3ForConditionalGeneration
from peft import PeftModel

BASE = "unsloth/gemma-3-4b-it"
LORA = "novanandriyono/javanese-qlora"

tok = AutoTokenizer.from_pretrained(LORA)
model = Gemma3ForConditionalGeneration.from_pretrained(
    BASE, torch_dtype=torch.bfloat16, device_map="auto"
)
model = PeftModel.from_pretrained(model, LORA)
model = model.merge_and_unload()
model.eval()

messages = [{"role": "user", "content": "Sugeng enjang! Critakna babagan Gunung Bromo."}]
inputs = tok.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt", return_dict=True
).to(model.device)

with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=512, do_sample=True,
                         temperature=0.7, top_p=0.9)

print(tok.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Option 2: 4-bit base with bitsandbytes (lower VRAM)

import torch
from transformers import AutoTokenizer, Gemma3ForConditionalGeneration, BitsAndBytesConfig
from peft import PeftModel

BASE = "unsloth/gemma-3-4b-it-unsloth-bnb-4bit"
LORA = "novanandriyono/javanese-qlora"

bnb = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
)

tok = AutoTokenizer.from_pretrained(LORA)
model = Gemma3ForConditionalGeneration.from_pretrained(
    BASE, quantization_config=bnb, device_map="auto"
)
model = PeftModel.from_pretrained(model, LORA)
model.eval()

If you get a key mismatch error when loading, try AutoModelForCausalLM instead of Gemma3ForConditionalGeneration.

Training details

Base model	unsloth/gemma-3-4b-it-unsloth-bnb-4bit
Method	QLoRA (PEFT 0.18.1, Unsloth)
Train as	Language model (text only)
Training data	TODO: describe the dataset (source, size, format)
LoRA rank / alpha	TODO
Epochs / steps	TODO
Learning rate	TODO
Max sequence length	TODO
Hardware	TODO

Limitations and bias

• The model can produce incorrect, inconsistent, or made-up content.
• Javanese has several speech levels (ngoko, madya, krama), and the model may mix them or pick the wrong register.
• Regional dialects and rare words may be poorly covered.
• It inherits the limitations and biases of the base model (Gemma 3) and of the training data.
• Vision input is not trained or evaluated in this adapter. Use text only.

Framework versions

• PEFT 0.18.1
• Transformers >= 4.50

Contact

Open a discussion on this repo's Community tab.