Hugging Face's logo

hkust-nlp
/
mstar-8b-v1.0

Safetensors
minicpmv
custom_code
Model card Files
xet
Community
1
PeterV09 commited on
Commit
4a7c02f
·
verified ·
1 Parent(s): 3f68cf1

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +102 -0
README.md ADDED
@@ -0,0 +1,102 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: mit
3
+ base_model:
4
+ - openbmb/MiniCPM-Llama3-V-2_5
5
+ ---
6
+
7
+ # M-STAR
8
+
9
+ <p align="center">
10
+ <img src="./assets/mstar-logo.png" width="500">
11
+ </p>
12
+
13
+ <p align="center">
14
+ <a href="https://mstar-lmm.github.io/">Project Page</a>&nbsp;&nbsp;&nbsp;
15
+ </p>
16
+
17
+ M-STAR is a framework to improve the **Multimodal Reasoning** ability of Large Multimodal Models (LMMs) via **Self-Evolving Training**.
18
+
19
+ Unlike traditional **Self-Evolving Training**, M-STAR supports **Large Multimodal Models**, **Training with Multimodal Process Reward Models (MPRM)**, and **Adaptive Explorations during Training**.
20
+
21
+ This is M-STAR-MiniCPM-V-2.5 model based on [MiniCPM-Llama3-V-2_5](https://huggingface.co/openbmb/MiniCPM-Llama3-V-2_5), which has been trained using M-STAR, our self-evolving training framework for multimodal reasoning.
22
+
23
+ - **M-STAR Resources**:
24
+
25
+ | **Component** |**Description** |
26
+ |------------------------------|---------------------------------------------------------------------------------------------------------------------|
27
+ | **M-STAR Model** | A strong LMM for multimodal reasoning, scoring **59.5** on MathVista, based on [MiniCPM-V-2.5](https://github.com/OpenBMB/MiniCPM-V) with 8B parameters. |
28
+ | **M-STAR PRM** | A Multimodal Process Reward Model (MPRM) that evaluates the quality of multimodal reasoning data at the step level. |
29
+ | **M-STAR CoT Dataset** | A collection of 100K generated multimodal reasoning data with CoT, where the queries are sourced from [MathV360K](https://huggingface.co/datasets/Zhiqiang007/MathV360K). |
30
+ | **M-STAR MPRM Training Dataset** | A set of 50K multimodal reasoning data designed for training MPRM. |
31
+
32
+
33
+ ## Performance
34
+
35
+ ### Main Results
36
+
37
+ <div align="center">
38
+
39
+ | | MathVista | FQA | GPS | MWP | TQA | VQA |
40
+ |----------------------------|-----------|-------|-------|-------|-------|-------|
41
+ | **Baselines** | | | | | | |
42
+ | MiniCPM-V-2.5 | 52.4 | 59.2 | 44.7 | 50.5 | 53.8 | 48.0 |
43
+ | &nbsp;&nbsp;&nbsp;+ warmup | 52.6 | 58.4 | 47.1 | 57.0 | 53.8 | 45.8 |
44
+ | SFT | 54.8 | 58.7 | 50.5 | 56.5 | 55.7 | 50.8 |
45
+ | ReST<sup>EM</sup> | 55.1 | 59.1 | 49.5 | 65.6 | 55.1 | 48.0 |
46
+ | Iterative RFT | 55.7 | 59.1 | 49.5 | 64.5 | 55.1 | 47.5 |
47
+ | **Static components only** | | | | | | |
48
+ | Cont. Self-Evolving | 57.2 | 57.6 | 56.3 | 65.1 | 57.0 | 49.7 |
49
+ | &nbsp;&nbsp;&nbsp;+ PRM Re-Rank | 59.2 | 59.1↑0.7 | 61.1↑14 | 68.3↑11.3 | 55.1↑1.3 | 51.4↑5.6 |
50
+ | **Automatically tuning the temperature T** | | | | | | |
51
+ | M-STAR (Reward-Pass@2) | 59.5 (+6.9) | 59.5↑1.1 | 59.1↑12 | 65.6↑8.6 | 58.9↑5.1 | 54.2↑8.4 |
52
+ | **Reference** | | | | | | |
53
+ | GPT-4o | 63.8 | - | - | - | - | - |
54
+ | Gemini 1.5 Flash | 58.4 | - | - | - | - | - |
55
+ | GPT-4T 2024-04-09 | 58.1 | - | - | - | - | - |
56
+ | Pixtral 12B | 58.0 | - | - | - | - | - |
57
+ | InternLM-XComposer2-VL-7B | 57.6 | 55.0 | 63.0 | 73.7 | 56.3 | 39.7 |
58
+ | Math-LLaVA-13B | 46.6 | 37.2 | 57.7 | 56.5 | 51.3 | 33.5 |
59
+ | LLaVA-NeXT-34B | 46.5 | - | - | - | - | - |
60
+
61
+ </div>
62
+
63
+ <div align="center">
64
+
65
+ | Model | MathVista | M3CoT | MMStar-R | MMBench-R | AI2D | Average |
66
+ |--------------------------|-----------|---------|----------|-----------|--------|----------|
67
+ | MiniCPM-V-2.5 | 52.4 | 41.2 | 44.6 | 72.6 | 64.4 | 55.0 |
68
+ | &nbsp;&nbsp;&nbsp;+ warmup | 52.6 | 47.8 | 45.1 | 76.9 | 65.9 | 57.7 |
69
+ | M-STAR | 59.5↑6.9 | 48.7↑0.9 | 50.7↑5.6 | 79.9↑3 | 69.1↑3.2 | 61.6↑3.9 |
70
+ | Phi-3.5-vision | 46.5 | 39.4 | 42.5 | 56.8 | 47.5 | 46.5 |
71
+ | &nbsp;&nbsp;&nbsp;+ warmup | 49.3 | 46.5 | 44.2 | 70.9 | 65.5 | 55.3 |
72
+ | M-STAR | 54.5↑5.2 | 51.3↑4.8 | 48.8↑4.6 | 73.6↑2.7 | 67.9↑2.4 | 59.2↑3.9 |
73
+ | InternVL2-2B | 46.4 | 16.7 | 20.0 | 14.2 | 33.5 | 26.2 |
74
+ | &nbsp;&nbsp;&nbsp;+ warmup | 47.6 | 45.6 | 41.8 | 68.8 | 60.0 | 52.8 |
75
+ | M-STAR | 50.3↑2.7 | 47.1↑1.5 | 42.0↑0.2 | 67.3↓1.5 | 59.7↓0.3 | 53.3↑0.5 |
76
+
77
+ </div>
78
+
79
+ ### Effectiveness of Adaptively Adjusting Exploration
80
+
81
+ <p align="center">
82
+ <img src="./assets/dynamic.png" width="500">
83
+ </p>
84
+
85
+ Evaluating the effectiveness of adaptively adjusting exploration:
86
+
87
+ - **Reward-Pass@2**: The percentage of samples for which there exist correct responses among the top 2 responses ranked by the reward model. This metric directly reflects the exploitation efficacy of the reward model for the current policy. We choose Pass@2 since our training strategy involves selecting the top 2 responses using the reward model.
88
+
89
+ "Static" refers to models trained without adaptive exploration, while "Dynamic" indicates those trained with this mechanism. All models shown were trained using the M-STAR framework with optimized components as explored in our paper.
90
+
91
+ ## M-STAR Resources
92
+ <div align="center">
93
+
94
+ | Resource | Link | License |
95
+ |------------------------------------------------|-----------|------------|
96
+ | **M-STAR Datasets**
97
+ | **M-STAR CoT Dataset** | ... | [MIT License](https://opensource.org/license/mit)
98
+ | **M-STAR MPRM Training Dataset** | ... | [MIT License](https://opensource.org/license/mit)
99
+ | **M-STAR Models** | | |
100
+ | M-STAR-8B-v1.0 | ... | [MiniCPM Model License](https://github.com/OpenBMB/MiniCPM/blob/main/MiniCPM%20Model%20License.md) |
101
+ | M-STAR-PRM-8B-v1.0 | ... | [MiniCPM Model License](https://github.com/OpenBMB/MiniCPM/blob/main/MiniCPM%20Model%20License.md) |
102
+ </div>