Towards Full Pipeline FP8 Reinforcement Learning for LLMs

Benchmark Model Rank Results
mathematical-reasoning-on-aime-2025Qwen3-8B-Base + Calibrated Clipping (Blockwise FP8)–Accuracy: 27.1
mathematical-reasoning-on-aime24Qwen3-8B-Base + Calibrated Clipping (Blockwise FP8)–Acc: 33.5
mathematical-reasoning-on-math-500Qwen3-8B-Base + Calibrated Clipping (Blockwise FP8)–Accuracy: 90.7
mathematical-reasoning-on-minerva-mathQwen3-8B-Base + Calibrated Clipping (Tensorwise FP8)–Accuracy: 44.6
mathematical-reasoning-on-olympiadbenchQwen3-8B-Base + Calibrated Clipping (Rowwise FP8)–Accuracy: 58.3