Datasets:
Update README.md
Browse files
README.md
CHANGED
|
@@ -33,8 +33,19 @@ configs:
|
|
| 33 |
|
| 34 |
[\[📂 GitHub\]](https://github.com/ModalMinds/MM-EUREKA) [\[📜 Paper\]](https://arxiv.org/abs/2503.07365v2)
|
| 35 |
|
|
|
|
|
|
|
| 36 |
***`2025/04/16:` We release a new version of MMK12, which can greatly enhance the multimodal reasoning of Qwen-2.5-VL.***
|
| 37 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 38 |
We use MMK12 for RL training to develop MM-EUREKA-7B and MM-EUREKA-32B, with specific training details available in [paper](https://arxiv.org/abs/2503.07365v2).
|
| 39 |
|
| 40 |
Both models demonstrate excellent performance on the MMK12 evaluation set (a multidisciplinary multimodal reasoning benchmark), with MM-EUREKA-32B ranking second only to o1.
|
|
|
|
| 33 |
|
| 34 |
[\[📂 GitHub\]](https://github.com/ModalMinds/MM-EUREKA) [\[📜 Paper\]](https://arxiv.org/abs/2503.07365v2)
|
| 35 |
|
| 36 |
+
|
| 37 |
+
***`2025/04/16:` MMK12 is a completely manually collected multimodal mathematical reasoning dataset. Compared to other current datasets, it can fully ensure the authenticity of answers, and all questions come from the real world, making it more diverse..***
|
| 38 |
***`2025/04/16:` We release a new version of MMK12, which can greatly enhance the multimodal reasoning of Qwen-2.5-VL.***
|
| 39 |
|
| 40 |
+
| | Scope | Type | Img. Source | QA Source | CoT Answer Source |
|
| 41 |
+
|---|---|---|---|---|---|
|
| 42 |
+
| MAVIS | Geo & Func | MCQ & FB | Synthetic | Synthetic Engine | GPT-4o |
|
| 43 |
+
| Geo3k | Geo | FB | Real world | Real world | None |
|
| 44 |
+
| RCOT | Geo | MCQ & FB | Synthetic | Synthetic Engine | GPT-4o |
|
| 45 |
+
| MultiMath | Diverse | MCQ & FB | Real World | GPT-4o | GPT-4o |
|
| 46 |
+
| MMK12 | Diverse | FB | Real World | Real World | Real World |
|
| 47 |
+
|
| 48 |
+
|
| 49 |
We use MMK12 for RL training to develop MM-EUREKA-7B and MM-EUREKA-32B, with specific training details available in [paper](https://arxiv.org/abs/2503.07365v2).
|
| 50 |
|
| 51 |
Both models demonstrate excellent performance on the MMK12 evaluation set (a multidisciplinary multimodal reasoning benchmark), with MM-EUREKA-32B ranking second only to o1.
|