Parameter	Value
Base model	Qwen/Qwen3-0.6B
Dataset	HuggingFaceH4/no_robots
Epochs	2
Batch size	32
Gradient accumulation	4
Learning rate	2e-5
Scheduler	cosine
Max sequence length	1024
Precision	bf16

Hardware

GPU: NVIDIA GB10 (Blackwell, sm_121)
Memory: 128GB unified CPU/GPU
CUDA: 13.0

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("natolambert/qwen3-dgx-spark-sft")
tokenizer = AutoTokenizer.from_pretrained("natolambert/qwen3-dgx-spark-sft")

messages = [{"role": "user", "content": "Write a short poem about machine learning."}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True)
outputs = model.generate(inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training Script

Trained using the DGX Spark SFT script from open-instruct:

./scripts/train/dgx-spark/sft_no_robots.sh

See dgx-spark-setup for details on running ML training on DGX Spark.

Limitations

This is primarily a test model to validate the DGX Spark training pipeline. It has not been extensively evaluated for downstream tasks.