Improve language tag
#2
by
lbourdois
- opened
README.md
CHANGED
|
@@ -1,91 +1,103 @@
|
|
| 1 |
-
---
|
| 2 |
-
base_model: Qwen/Qwen2.5-32B-Instruct
|
| 3 |
-
language:
|
| 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 |
-
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
|
|
| 68 |
-
|
| 69 |
-
**
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
|
| 90 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 91 |
```
|
|
|
|
| 1 |
+
---
|
| 2 |
+
base_model: Qwen/Qwen2.5-32B-Instruct
|
| 3 |
+
language:
|
| 4 |
+
- zho
|
| 5 |
+
- eng
|
| 6 |
+
- fra
|
| 7 |
+
- spa
|
| 8 |
+
- por
|
| 9 |
+
- deu
|
| 10 |
+
- ita
|
| 11 |
+
- rus
|
| 12 |
+
- jpn
|
| 13 |
+
- kor
|
| 14 |
+
- vie
|
| 15 |
+
- tha
|
| 16 |
+
- ara
|
| 17 |
+
library_name: transformers
|
| 18 |
+
license: apache-2.0
|
| 19 |
+
tags:
|
| 20 |
+
- multi-agent systems
|
| 21 |
+
- multiagent-collaboration
|
| 22 |
+
- reasoning
|
| 23 |
+
- mathematics
|
| 24 |
+
- code
|
| 25 |
+
pipeline_tag: text-generation
|
| 26 |
+
model-index:
|
| 27 |
+
- name: m1-32b
|
| 28 |
+
results: []
|
| 29 |
+
---
|
| 30 |
+
|
| 31 |
+
[Two Heads are Better Than One: Test-time Scaling of Multi-agent Collaborative Reasoning](https://arxiv.org/pdf/2504.09772)
|
| 32 |
+
|
| 33 |
+
**M1-32B** is a 32B-parameter large language model fine-tuned from [Qwen2.5-32B-Instruct](https://arxiv.org/pdf/2412.15115) on the **M500** dataset—an interdisciplinary multi-agent collaborative reasoning dataset. M1-32B is optimized for improved reasoning, discussion, and decision-making in multi-agent systems (MAS), including frameworks such as [AgentVerse](https://github.com/OpenBMB/AgentVerse).
|
| 34 |
+
|
| 35 |
+
Code: [https://github.com/jincan333/MAS-TTS](https://github.com/jincan333/MAS-TTS)
|
| 36 |
+
|
| 37 |
+
---
|
| 38 |
+
|
| 39 |
+
## 🚀 Key Features
|
| 40 |
+
|
| 41 |
+
- 🧠 **Enhanced Collaborative Reasoning**
|
| 42 |
+
Trained on real multi-agent traces involving diverse roles like Expert Recruiter, Problem Solvers, and Evaluator.
|
| 43 |
+
|
| 44 |
+
- 🗣️ **Role-Aware Dialogue Generation**
|
| 45 |
+
Learns to reason and respond from different expert perspectives based on structured prompts.
|
| 46 |
+
|
| 47 |
+
- ⚙️ **Optimized for Multi-Agent Systems**
|
| 48 |
+
Performs well as a MAS agent with adaptive collaboration and token budgeting.
|
| 49 |
+
|
| 50 |
+
---
|
| 51 |
+
|
| 52 |
+
## 🏗️ Model Training
|
| 53 |
+
|
| 54 |
+
- **Base Model:** Qwen2.5-32B-Instruct
|
| 55 |
+
- **Dataset:** [M500](https://huggingface.co/datasets/Can111/M500) (500 curated multi-agent reasoning traces)
|
| 56 |
+
- **Objective:** Supervised Fine-Tuning (SFT) on role-conditioned prompts
|
| 57 |
+
- **Training Setup:**
|
| 58 |
+
- 8 × A100 GPUs
|
| 59 |
+
- 5 epochs
|
| 60 |
+
- Learning rate: 1e-5
|
| 61 |
+
- Frameworks: DeepSpeed, FlashAttention, LLaMA-Factory
|
| 62 |
+
|
| 63 |
+
---
|
| 64 |
+
|
| 65 |
+
## 📊 Performance
|
| 66 |
+
|
| 67 |
+
| **Model** | **General Understanding** | | **Mathematical Reasoning** | | **Coding** | |
|
| 68 |
+
|--------------------------|---------------------------|----------------|-----------------------------|------------|----------------|-----------|
|
| 69 |
+
| | **GPQA** | **Commongen** | **AIME2024** | **MATH-500** | **HumanEval** | **MBPP-S**|
|
| 70 |
+
| **Non-Reasoning Models** | | | | | | |
|
| 71 |
+
| Qwen2.5 | 50.2 | 96.7 | 21.1 | 84.4 | 89.0 | 80.2 |
|
| 72 |
+
| DeepSeek-V3 | **58.6** | **98.6** | **33.3** | **88.6** | 89.6 | 83.9 |
|
| 73 |
+
| GPT-4o | 49.2 | 97.8 | 7.8 | 81.3 | **90.9** | **85.4** |
|
| 74 |
+
| **Reasoning Models** | | | | | | |
|
| 75 |
+
| s1.1-32B | 58.3 | 94.1 | 53.3 | 90.6 | 82.3 | 77.4 |
|
| 76 |
+
| DeepSeek-R1 | **75.5** | 97.2 | 78.9 | **96.2** | **98.2** | 91.7 |
|
| 77 |
+
| o3-mini | 71.3 | **99.1** | **84.4** | 95.3 | 97.0 | **93.6** |
|
| 78 |
+
| M1-32B (Ours) | 61.1 | 96.9 | 60.0 | 95.1 | 92.8 | 89.1 |
|
| 79 |
+
| M1-32B w. CEO (Ours) | 62.1 | 97.4 | 62.2 | 95.8 | 93.9 | 90.5 |
|
| 80 |
+
|
| 81 |
+
**Table Caption:**
|
| 82 |
+
Performance comparison on general understanding, mathematical reasoning, and coding tasks using strong reasoning and non-reasoning models within the AgentVerse framework. Our method achieves substantial improvements over Qwen2.5 and s1.1-32B on all tasks, and attains performance comparable to o3-mini and DeepSeek-R1 on MATH-500 and MBPP-S, demonstrating its effectiveness in enhancing collaborative reasoning in MAS. Note that the results of s1.1-32B are obtained without using budget forcing.
|
| 83 |
+
|
| 84 |
+
---
|
| 85 |
+
|
| 86 |
+
## 💬 Intended Use
|
| 87 |
+
|
| 88 |
+
M1-32B is intended for research on Multi-agent reasoning and collaboration in MAS
|
| 89 |
+
|
| 90 |
+
---
|
| 91 |
+
|
| 92 |
+
## Citation
|
| 93 |
+
|
| 94 |
+
If you use this model, please cite the relevant papers:
|
| 95 |
+
|
| 96 |
+
```bibtex
|
| 97 |
+
@article{jin2025two,
|
| 98 |
+
title={Two Heads are Better Than One: Test-time Scaling of Multi-agent Collaborative Reasoning},
|
| 99 |
+
author={Jin, Can and Peng, Hongwu and Zhang, Qixin and Tang, Yujin and Metaxas, Dimitris N and Che, Tong},
|
| 100 |
+
journal={arXiv preprint arXiv:2504.09772},
|
| 101 |
+
year={2025}
|
| 102 |
+
}
|
| 103 |
```
|