Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,34 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
---
|
| 4 |
+
|
| 5 |
+
# Kaggle AI Mathematical Olympiad - Progress Prize 2 - 9th Place Solution (Fast-Math-R1-14B)
|
| 6 |
+
## Team
|
| 7 |
+
- Hiroshi Yoshihara @ [Aillis Inc.](https://aillis.jp/en), [The Univ. of Tokyo](https://publichealth.f.u-tokyo.ac.jp/#page_home)
|
| 8 |
+
- Yuichi Inoue @ [Sakana AI](https://sakana.ai)
|
| 9 |
+
- Taiki Yamaguchi @ [Rist Inc.](https://www.rist.co.jp/en/)
|
| 10 |
+
|
| 11 |
+
# Summary
|
| 12 |
+
By applying SFT and GRPO on difficult math problems, we enhanced the performance of `DeepSeek-R1-Distill-Qwen-14B` and developed `Fast-Math-R1-14B`,
|
| 13 |
+
which achieves up to 60% faster inference while maintaining accuracy.
|
| 14 |
+
|
| 15 |
+
Technical details can be found in [Kaggle Discussion](https://www.kaggle.com/competitions/ai-mathematical-olympiad-progress-prize-2/discussion/571252) and [Github](https://github.com/analokmaus/kaggle-aimo2-fast-math-r1).
|
| 16 |
+
|
| 17 |
+
<img src="https://github.com/analokmaus/kaggle-aimo2-fast-math-r1/blob/8eb55edfdb8e922b2d504000fb1cefe22acf67ef/assets/pass1_aime2024.png?raw=true" width="50%"><img src="https://github.com/analokmaus/kaggle-aimo2-fast-math-r1/blob/8eb55edfdb8e922b2d504000fb1cefe22acf67ef/assets/pass1_aime2025.png?raw=true" width="50%">
|
| 18 |
+
| | | AIME 2024 | | AIME 2025 | |
|
| 19 |
+
| ---------------------------- | ------------ | ---------------- | ------------- | ---------------- | ------------- |
|
| 20 |
+
| Model | Token budget | Pass@1 (avg. 64) | Output tokens | Pass@1 (avg. 64) | Output tokens |
|
| 21 |
+
| DeepSeek-R1-Distill-Qwen-14B | 16384 | 63.3 | 9590 | 46.7 | 10602 |
|
| 22 |
+
| | 12800 | 58 | 6444 | 41.9 | 6684 |
|
| 23 |
+
| | 8192 | 45.6 | 4920 | 30.6 | 4611 |
|
| 24 |
+
| Light-R1-14B-DS | 16384 | **66.8** | 10146 | **51.3** | 11308 |
|
| 25 |
+
| | 12800 | 59.2 | 6974 | 43.8 | 6869 |
|
| 26 |
+
| | 8192 | 42.4 | 5500 | 30.4 | 4908 |
|
| 27 |
+
| Fast-Math-R1-14B | 16384 | 66 | **7932** | 49.2 | **9066** |
|
| 28 |
+
| | 12800 | **63** | **5996** | **46.1** | **6127** |
|
| 29 |
+
| | 8192 | **51.4** | **4269** | **37.2** | **3905** |
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
# Dataset
|
| 33 |
+
- [Our first stage SFT dataset](https://huggingface.co/datasets/RabotniKuma/Fast-Math-R1-SFT)
|
| 34 |
+
- [Our second stage GRPO dataset](https://huggingface.co/datasets/RabotniKuma/Fast-Math-R1-GRPO)
|