CoolFace
Modelpublic

Nickyang/FastCuRL-1.5B-V2

sourceHugging Facemitupdated 1y agoView on Hugging Face
1likes16downloads
Model Card

<div align="center"> <span style="font-family: default; font-size: 1.5em;">FastCuRL-1.5B-V2</span> </div>

FastCuRL Overview

2025-05-23

We release FastCuRL-1.5B-V3 and FastCuRL-1.5B-V2.

2025-03-17

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

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

2025-03-21

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

Key Results

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

ModelAIME 2024MATH 500AMC 2023Minerva MathOlympiadBenchAvg.
Qwen2.5-Math-7B-Instruct13.379.850.634.640.743.8
rStar-Math-7B26.778.447.5-47.1-
Eurus-2-7B-PRIME26.779.257.838.642.148.9
Qwen2.5-7B-SimpleRL26.782.462.5<strong>39.7</strong>43.350.9
DeepSeek-R1-Distill-Qwen-1.5B28.882.862.926.543.348.9
Still-1.5B32.584.466.729.045.451.6
DeepScaleR-1.5B-Preview43.187.873.630.250.057.0
<strong>FastCuRL-1.5B-Preview</strong>43.188.074.231.650.457.5
<strong>FastCuRL-1.5B-V2</strong>47.589.377.032.853.360.0
<strong>FastCuRL-1.5B-V3</strong><strong>49.6</strong><strong>90.5</strong><strong>78.5</strong><strong>34.7</strong><strong>54.5</strong><strong>61.6</strong>

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

Acknowledgements