Fastcurl 1.5B Preview

🚀 FastCuRL-1.5B-Preview

FastCuRL-1.5B-Preview is a slow - thinking reasoning model that outperforms the previous state - of - the - art model with only 50% of the training steps. The code, model, and data are open - sourced for better reproducibility and research advancement.

🚀 Quick Start

The key milestones of FastCuRL-1.5B-Preview are as follows:

2025-03-17

We released FastCuRL-1.5B-Preview, a slow - thinking reasoning model that outperforms the previous SoTA DeepScaleR-1.5B-Preview with 50% training steps! We adapted a novel curriculum - guided iterative lengthening reinforcement learning to the DeepSeek-R1-Distill-Qwen-1.5B and observed continuous performance improvement as training steps increased. To better reproduce our work and advance research progress, we open - sourced our code, model, and data.

Code: https://github.com/nick7nlp/FastCuRL

2025-03-21

Paper: https://arxiv.org/abs/2503.17287

✨ Features

Key Results

We report Pass@1 accuracy averaged over 16 samples for each problem.

Training Data

Following DeepScaleR, our training dataset consists of 40,315 unique problem - answer pairs compiled from:

• AIME problems (1984 - 2023)
• AMC problems (before 2023)
• Omni - MATH dataset
• Still dataset

📚 Documentation

Acknowledgements

• Our training experiments are powered by our heavily modified fork of verl and deepscaler.
• Our model is trained on top of DeepSeek-R1-Distill-Qwen-1.5B.

📄 License

This project is under the MIT license.