Instructions to use Balab2021/Nemotron-3.5-Lightning-30B-A3B-Olympics-Reasoning with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Balab2021/Nemotron-3.5-Lightning-30B-A3B-Olympics-Reasoning with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="Balab2021/Nemotron-3.5-Lightning-30B-A3B-Olympics-Reasoning")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Balab2021/Nemotron-3.5-Lightning-30B-A3B-Olympics-Reasoning")
model = AutoModelForCausalLM.from_pretrained("Balab2021/Nemotron-3.5-Lightning-30B-A3B-Olympics-Reasoning", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Balab2021/Nemotron-3.5-Lightning-30B-A3B-Olympics-Reasoning with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Balab2021/Nemotron-3.5-Lightning-30B-A3B-Olympics-Reasoning"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Balab2021/Nemotron-3.5-Lightning-30B-A3B-Olympics-Reasoning",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Balab2021/Nemotron-3.5-Lightning-30B-A3B-Olympics-Reasoning

SGLang

How to use Balab2021/Nemotron-3.5-Lightning-30B-A3B-Olympics-Reasoning with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Balab2021/Nemotron-3.5-Lightning-30B-A3B-Olympics-Reasoning" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Balab2021/Nemotron-3.5-Lightning-30B-A3B-Olympics-Reasoning",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Balab2021/Nemotron-3.5-Lightning-30B-A3B-Olympics-Reasoning" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Balab2021/Nemotron-3.5-Lightning-30B-A3B-Olympics-Reasoning",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Balab2021/Nemotron-3.5-Lightning-30B-A3B-Olympics-Reasoning with Docker Model Runner:
```
docker model run hf.co/Balab2021/Nemotron-3.5-Lightning-30B-A3B-Olympics-Reasoning
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

Nemotron-3.5-Lightning-30B-A3B-Olympics-Reasoning

Full-parameter SFT of nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 on Balab2021/Olympics-Reasoning-Nemotron-3.5-Lightning: verified step-by-step reasoning traces over tables from the 120-year Olympics athlete-events data, one per source row. The traces were distilled from the base model itself, served with NVIDIA Dynamo (SGLang backend), keeping the shortest trace whose answer matched the exact gold answer (rejection-sampling self-distillation).

Trained with Megatron-Bridge (NeMo 26.08.01) on NVIDIA GB200 NVL72 nodes; the assistant turn is trained in Nemotron's native reasoning format (<think>...</think>, then Final answer: <answer>).

setting / result	value
training data	Balab2021/Olympics-Reasoning-Nemotron-3.5-Lightning train split: 227,051 rows (snapshot at 99.575% source-row coverage), 1 epoch
hardware	NVIDIA GB200 NVL72 (Ptyche), Megatron-Bridge, HybridEP MoE dispatcher
row benchmark (23,861 q, held-out athletes): pass@1	92.0% -> 92.35%
row benchmark (23,861 q, held-out athletes): mean tokens	2,209 -> 1,951
row benchmark (23,861 q, held-out athletes): truncated	3.02% -> 2.23%
table benchmark (550 q): pass@1	93.98% -> 93.05%
table benchmark (550 q): mean tokens	2,467 -> 2,529
table benchmark (550 q): truncated	4.59% -> 5.32%

Usage

Same as the base model (vLLM / SGLang / Dynamo / Transformers, enable_thinking=True, T=1.0, top_p=0.95). The training system prompt: "You are a sports data analyst. Use only the table you are given. Work through the problem step by step, then give the final answer on its own line as 'Final answer: '."

Downloads last month: 28

Safetensors

Model size

32B params

Tensor type

BF16

Model tree for Balab2021/Nemotron-3.5-Lightning-30B-A3B-Olympics-Reasoning

Base model

nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16

Finetuned

(29)

this model

Quantizations

1 model

Balab2021
/

Nemotron-3.5-Lightning-30B-A3B-Olympics-Reasoning

Nemotron-3.5-Lightning-30B-A3B-Olympics-Reasoning

Usage

Model tree for Balab2021/Nemotron-3.5-Lightning-30B-A3B-Olympics-Reasoning

Dataset used to train Balab2021/Nemotron-3.5-Lightning-30B-A3B-Olympics-Reasoning