richardcsuwandi/nemotron3-nusantara-lora

Nemotron-3-Super Indonesian and Nusantara LoRA (AutoScientist)

A small LoRA adapter (rank 8, q_proj and v_proj, about 3 MB) for nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16, fine-tuned with Adaption AutoScientist to be better at Indonesian and at translating between Indonesian, English and Indonesia's regional languages.

Training data (about 19K rows, prompt/completion)

• NusaX: parallel sentences between Indonesian and ten regional languages (Javanese, Sundanese, Minangkabau, Balinese, Buginese, Acehnese, Banjarese, Madurese, Ngaju, Batak Toba).
• FLORES-200: English to eight of those languages.

Results

AutoScientist held-out domain evaluation (language domain, judged against the base model, about 96 prompts): this adapter won 95.3% of comparisons. Replicate runs of the same recipe scored between 92.6% and 97.4%, so treat a single number as noisy (about 3 points of spread). This is a pairwise win rate against the base model, not a benchmark score.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, device_map="auto", torch_dtype="auto")
model = PeftModel.from_pretrained(model, "richardcsuwandi/nemotron3-nusantara-lora")

Limitations

Short-sentence translation data, so it is weaker for long documents and formal registers. Regional-language quality varies with how much the base model already knew. Licenses: follow the base model's license, and the CC BY-SA terms of NusaX and FLORES-200.