1.5B and 7B models training on 3 NP Graph Problems and achieved SOTA as GPT-4o and QwQ-32B