CompactAI/LFM2.5-1.2B-Thinking-math-safe at main