File size: 3,762 Bytes
0e1aaee
 
 
 
5ed0208
0e1aaee
 
 
 
 
5ed0208
0e1aaee
5ed0208
0e1aaee
 
03f3b95
 
0e1aaee
 
5ed0208
0e1aaee
5ed0208
 
0e1aaee
 
 
4936a51
 
 
5ed0208
 
0e1aaee
 
 
 
4936a51
7d77b92
 
4936a51
0e1aaee
 
4936a51
0e1aaee
 
 
 
5ed0208
 
0e1aaee
5ed0208
 
 
0e1aaee
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
# Online-DPO-R1
* **Blog**: https://www.notion.so/Online-DPO-R1-1908b9a70e7b80c3bc83f4cf04b2f175
* **Authors**: 
* **Code**: https://github.com/RLHFlow/Online-DPO-R1

## Introduction
We release unofficial checkpoints for PPO, iterative DPO and rejection sampling (RAFT) trained from Qwen2.5-MATH-7B-base with rule-based RL, which are based on the success of Deepseek-R1-Zero and recent replications of PPO approach.
Evaluated on five widely-adopted benchmarks **AIME 2024**, **MATH 500**, **AMC**, **Minerva Math**, **OlympiadBench**, our **iterative DPO** and **RAFT** model achieve 
significant enhancement compared to the base model and are comparable to the PPO approach.
Our models are trained by using the prompt set from the MATH training set and Numina Math.

Moreover, we provide a [detailed recipe](https://github.com/RLHFlow/Online-DPO-R1) to reproduce the model. Enjoy!

## Model Releases
- [PPO model] (https://huggingface.co/RLHFlow/Qwen2.5-7B-PPO-Zero)
- [Iterative DPO from SFT model] (https://huggingface.co/RLHFlow/Qwen2.5-7B-DPO)
- [Iterative DPO from base model] (https://huggingface.co/RLHFlow/Qwen2.5-7B-DPO-Zero)
- [Iterative DPO with Negative Log-Likelihood (NLL)] (https://huggingface.co/RLHFlow/Qwen2.5-7B-DPO-NLL-Zero)
- [Raft] (https://huggingface.co/RLHFlow/Qwen2.5-7B-RAFT-Zero)

## Dataset 


## Training methods
- Iterative DPO: Following the RLHF Workflow framework (https://arxiv.org/pdf/2405.07863), in each iteration, we sample multiple responses from the last trained policy, rank them via the ruled-based reward, and construct the preference pairs.
  Then, we optimize the policy by minimizing the DPO loss and enter the next iteration.
  Online iterative DPO can mitigate the issue of distribution shift and the limited coverage of offline data effectively

More detailed can be found in our [blog](https://www.notion.so/Online-DPO-R1-1908b9a70e7b80c3bc83f4cf04b2f175)! 


## Performance
| **Model**                  | **AIME 2024** | **MATH 500** | **AMC** | **Minerva Math** | **OlympiadBench** | **Average** |
|----------------------------|---------------|--------------|---------|------------------|-------------------|-------------|
| **Ours**                   |            |           |            |          |               |              |
| RLHFlow/Qwen2.5-7B-PPO-Zero        | **43.3 (+26.6)**   | 79.4 (+27.0)      | **62.5 (+10.0)**      | 33.1   (+20.2)      | 40.7   (+24.3)     | **51.8 (+21.6)** |
| RLHFlow/Qwen2.5-7B-DPO-Zero        | 26.7 (+10.0)       | 76.8 (+24.4)      | **62.5 (+10.0)**      | 30.9 (+18.0)        | 37.9 (+21.5)       | 47.0 (+16.8)     |
| RLHFlow/Qwen2.5-7B-DPO             | 30.0 (+13.3)       | **84.4 (+32.0)**  | **62.5 (+10.0)**      | **33.5 (+20.6)**    | **48.4 (+32.0)**   | **51.8 (+21.6)** |
| RLHFlow/Qwen2.5-7B-RAFT-Zero       | 20.0 (+3.3)        | 77.6 (+25.2)      | 55.0 (+2.5)           | 30.5   (+17.6)      | 38.7   (+22.3)     | 44.4   (+14.2)   |
| **Baselines**             |            |           |            |          |               |              |
| Qwen2.5-Math-7B-Base       | 16.7       | 52.4      | 52.5       | 12.9     | 16.4          | 30.2         |
| Qwen2.5-Math-7B-Base + SFT Warm-up | 20.0       | 73.2      | 62.5       | 30.5     | 35.6          | 44.4         |
| Qwen-2.5-Math-7B-Instruct  | 13.3       | 79.8      | 50.6       | 34.6     | 40.7          | 43.8         |
| Llama-3.1-70B-Instruct     | 16.7       | 64.6      | 30.1       | 35.3     | 31.9          | 35.7         |
| Eurus-2-7B-PRIME           | 26.7       | 79.2      | 57.8       | 38.6     | 42.1          | 48.9         |
| GPT-4o                     | 9.3        | 76.4      | 45.8       | 36.8     | 43.3          | 43.3         |


## Usage



## Citation