DeepScaleR-1.5B-Preview 是一個語言模型，它基於 DeepSeek-R1-Distilled-Qwen-1.5B 進行微調，並使用分散式強化學習 (RL) 來擴展到長上下文長度。該模型在 AIME 2024 上達到 43.1% 的 Pass@1 準確度，相較於基礎模型 (28.8%) 提升了 15%，並僅以 1.5B 參數超越了 OpenAI 的 O1-Preview 效能。

模型	AIME 2024	MATH 500	AMC 2023	Minerva Math	Olympiad Bench	平均
DeepScaleR-1.5B-Preview	43.1	87.8	73.6	30.2	50.0	57.0
DeepSeek-R1-Distill-Qwen-1.5B	28.8	82.8	62.9	26.5	43.3	48.9
O1-Preview	40.0	81.4	-	-	-	-

資料

我們的訓練資料集包含約 40,000 個獨特的問題解答對，編譯自

AIME 問題 (1984-2023)
AMC 問題 (2023 年之前)
Omni-MATH 資料集
Still 資料集

評估

我們在競賽級數學基準測試中評估我們的模型，包括 AIME 2024、AMC 2023、MATH-500、Minerva Math 和 OlympiadBench。以下報告 Pass@1 準確度，每個問題平均取樣 16 次。

模型	AIME 2024	MATH 500	AMC 2023	Minerva Math	OlympiadBench	平均
Qwen-2.5-Math-7B-Instruct	13.3	79.8	50.6	34.6	40.7	43.8
rStar-Math-7B	26.7	78.4	47.5	-	47.1	-
Eurus-2-7B-PRIME	26.7	79.2	57.8	38.6	42.1	48.9
Qwen2.5-7B-SimpleRL	26.7	82.4	62.5	39.7	43.3	50.9
DeepSeek-R1-Distill-Qwen-1.5B	28.8	82.8	62.9	26.5	43.3	48.9
Still-1.5B	32.5	84.4	66.7	29.0	45.4	51.6
DeepScaleR-1.5B-Preview	43.1	87.8	73.6	30.2	50.0	57.0
O1-Preview	40.0	81.4	-	-	-	-

我們將 DeepScaleR 與我們使用的基礎 DeepSeek 模型，以及近期探索推理任務強化學習的學術著作進行比較。DeepScaleR 在所有基準測試中顯著優於基礎模型，在 AIME2024 上取得了 14.4% 的絕對提升，整體改進了 8.1%。此外，DeepScaleR 超越了近期的學術著作，如 rSTAR、Prime 和 SimpleRL，這些著作都是從 7B 模型微調而來。DeepScaleR 僅以 1.5B 參數就達到了 O1-preview 等級的效能，展現了卓越的效率提升。

參考文獻

文章

GitHub

Hugging Face

# DeepScaleR

🚀 Democratizing Reinforcement Learning for LLMs 🌟

DeepScaleR-1.5B-Preview is a language model fine-tuned from DeepSeek-R1-Distilled-Qwen-1.5B using distributed reinforcement learning (RL) to scale up to long context lengths. The model achieves 43.1% Pass@1 accuracy on AIME 2024, representing a 15% improvement over the base model (28.8%) and surpassing OpenAI's O1-Preview performance with just 1.5B parameters.

| Model | AIME 2024 | MATH 500 | AMC 2023 | Minerva Math | Olympiad Bench | Avg. |
| --- | --- | --- | --- | --- | --- | --- |
| **DeepScaleR-1.5B-Preview** | **43.1** | **87.8** | **73.6** | **30.2** | **50.0** | **57.0** |
| DeepSeek-R1-Distill-Qwen-1.5B | 28.8 | 82.8 | 62.9 | 26.5 | 43.3 | 48.9 |
| O1-Preview | 40.0 | 81.4 | - | - | - | - |

![image.png](/assets/library/deepscaler/c8dcb4f4-481a-4fea-89a9-168cdd1f48fe)

### Data
Our training dataset consists of approximately 40,000 unique problem-answer pairs compiled from:

- AIME problems (1984-2023)
- AMC problems (prior to 2023)
- Omni-MATH dataset
- Still dataset

### Evaluation

We evaluate our model on competition-level mathematics benchmarks, including AIME 2024, AMC 2023, MATH-500, Minerva Math, and OlympiadBench. Below, Pass@1 accuracy is reported, averaged over 16 samples for each problem.

| **Model** | **AIME 2024** | **MATH 500** | **AMC 2023** | **Minerva Math** | **OlympiadBench** | **Avg.** |
| --- | --- | --- | --- | --- | --- | --- |
| Qwen-2.5-Math-7B-Instruct | 13.3 | 79.8 | 50.6 | 34.6 | 40.7 | 43.8 |
| rStar-Math-7B | 26.7 | 78.4 | 47.5 | - | 47.1 | - |
| Eurus-2-7B-PRIME | 26.7 | 79.2 | 57.8 | 38.6 | 42.1 | 48.9 |
| Qwen2.5-7B-SimpleRL | 26.7 | 82.4 | 62.5 | **39.7** | 43.3 | 50.9 |
| DeepSeek-R1-Distill-Qwen-1.5B | 28.8 | 82.8 | 62.9 | 26.5 | 43.3 | 48.9 |
| Still-1.5B | 32.5 | 84.4 | 66.7 | 29.0 | 45.4 | 51.6 |
| **DeepScaleR-1.5B-Preview** | **43.1** | **87.8** | **73.6** | 30.2 | **50.0** | **57.0** |
| O1-Preview | 40.0 | 81.4 | - | - | - | - |

We compare DeepScaleR with the base DeepSeek model we use, as well as recent academic works exploring RL for reasoning tasks. DeepScaleR significantly outperforms the base model across all benchmarks, achieving a 14.4% absolute gain on AIME2024 and an 8.1% overall improvement. Additionally, DeepScaleR surpasses recent academic works such as rSTAR, Prime, and SimpleRL, which are finetuned from 7B models. DeepScaleR achieves O1-preview-level performance with only 1.5B parameters—a remarkable efficiency gain.

![image.png](/assets/library/deepscaler/afbe7f8e-04db-4350-ab39-352c023a6218)

## References

[Article](https://pretty-radio-b75.notion.site/DeepScaleR-Surpassing-O1-Preview-with-a-1-5B-Model-by-Scaling-RL-19681902c1468005bed8ca303013a4e2)

[GitHub](https://github.com/agentica-project/deepscaler)

[Hugging Face](https://huggingface.co/agentica-org/DeepScaleR-1.5B-Preview)

貼上、拖曳或點擊以上傳圖片 (.png, .jpeg, .jpg, .svg, .gif)