DeepScaleR

lastmod 2025-02-13created 2025-02-131 edits

Surpassing O1-Preview with a 1.5B model using RL

About

• DeepScaleR-1.5B-Preview: a model fine-tuned from Deepseek-R1-Distilled-Qwen-1.5B with reinforcement learning (RL)
• Achieves 43.1% AIME2024 Pass@1 accuracy (+14.3% over the base model),
  • Surpasses OpenAI o1-preview performance!
• Trained in 3,800 A100 GPU hours ($4500) → RL scaling 18.42x more efficient than 70,000 A100 GPU hours
• Dataset, code, and training logs released as open source → anyone can experiment with scaling intelligence using RL

See also

• DeepSeek

Favorite site

• DeepScaleR: Surpassing O1-Preview with a 1.5B Model by Scaling RL
• DeepScaleR: Surpassing O1-Preview with a 1.5B model using RL | GeekNews