Supervised fine-tune (full SFT) of nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 on the nvidia/Nemotron-SFT-Safety-v2 dataset, aimed at improving safe-response behavior while preserving the base model's reasoning ability.

Model details

Base model: nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 (30B total / ~3B active, MoE + Mamba hybrid)
Fine-tuning method: Full supervised fine-tuning (all parameters)
Dataset: nvidia/Nemotron-SFT-Safety-v2 (English subset)
Precision: BF16
Framework: Megatron-Bridge

Training configuration

Setting	Value
Parallelism	Tensor (TP) = 1, Expert (EP) = 8
Optimizer	Adam (fp32 states, CPU offload enabled)
Hardware	8x GPU node
Reasoning format	`<think> ... </think>` traces preserved from the dataset
The dataset was preprocessed into chat-templated `input`/`output` pairs using the
base model's tokenizer, with reasoning content wrapped in `<think>...</think>`
ahead of the final response.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "Pongsasit/nemotron-3-sfted"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id, trust_remote_code=True, torch_dtype="bfloat16", device_map="auto"
)
messages = [{"role": "user", "content": "How do I keep my online accounts secure?"}]
inputs = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt"
).to(model.device)
outputs = model.generate(inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))