Update README.md
Browse files
README.md
CHANGED
|
@@ -90,23 +90,22 @@ The plot below highlights the alignment comparison of the model trained with Con
|
|
| 90 |
### Benchmark Results Table
|
| 91 |
The table below summarizes the evaluation results across mathematical tasks and original capabilities for various models and training approaches.
|
| 92 |
|
| 93 |
-
| **Model** | **Math Tasks**
|
| 94 |
-
|
| 95 |
-
| | **MathHard**
|
| 96 |
-
|
| 97 |
-
| Llama3.1-8B-Instruct | 23.7
|
| 98 |
-
| OpenMath2-Llama3.1 | 38.4
|
| 99 |
-
| **Full Param Tune** | **38.5**
|
| 100 |
-
| Partial Param Tune | 36.4
|
| 101 |
-
| Stack Expansion | 35.6
|
| 102 |
-
| Hybrid Expansion | 34.4
|
| 103 |
-
| **Control LLM*** | 38.1
|
|
|
|
|
|
|
| 104 |
|
| 105 |
### Explanation of Groups
|
| 106 |
- **Math Tasks**:
|
| 107 |
- Covers **MathHard**, **Math**, and **GSM8K**, measuring the model's performance on mathematical reasoning and problem-solving tasks.
|
| 108 |
- **Original Capabilities**:
|
| 109 |
- Includes **ARC**, **GPQA**, **MMLU**, and **MMLUP**, reflecting the model’s ability to handle general reasoning and knowledge benchmarks.
|
| 110 |
-
|
| 111 |
-
|
| 112 |
-
|
|
|
|
| 90 |
### Benchmark Results Table
|
| 91 |
The table below summarizes the evaluation results across mathematical tasks and original capabilities for various models and training approaches.
|
| 92 |
|
| 93 |
+
| **Model** | **Math Tasks** | | | | **Original Capabilities** | | | | **Overall Avg.** |
|
| 94 |
+
|--------------------------|----------------------------|----------|-----------|----------|-----------------------------|---------|---------|-----------|------------------|
|
| 95 |
+
| | **MathHard** | **Math** | **GSM8K** | **Avg.** | **ARC** | **GPQA**| **MMLU**| **MMLUP** | |
|
| 96 |
+
|--------------------------|----------------------------|----------|-----------|----------|-----------------------------|---------|---------|-----------|------------------|
|
| 97 |
+
| Llama3.1-8B-Instruct | 23.7 | 50.9 | 85.6 | 52.1 | 83.4 | 29.9 | 72.4 | 46.7 | 56.3 |
|
| 98 |
+
| OpenMath2-Llama3.1 | 38.4 | 64.1 | 90.3 | 64.3 | 45.8 | 1.3 | 4.5 | 19.5 | 38.6 |
|
| 99 |
+
| **Full Param Tune** | **38.5** | **63.7** | 90.2 | **63.9** | 58.2 | 1.1 | 7.3 | 23.5 | 40.1 |
|
| 100 |
+
| Partial Param Tune | 36.4 | 61.4 | 89.0 | 61.8 | 66.2 | 6.0 | 25.7 | 30.9 | 45.6 |
|
| 101 |
+
| Stack Expansion | 35.6 | 61.0 | 90.8 | 61.8 | 69.3 | 18.8 | 61.8 | 43.1 | 57.6 |
|
| 102 |
+
| Hybrid Expansion | 34.4 | 61.1 | 90.1 | 61.5 | **81.8** | **25.9**| 67.2 | **43.9** | 59.3 |
|
| 103 |
+
| **Control LLM*** | 38.1 | 62.7 | **90.4** | 63.2 | 79.7 | 25.2 | **68.1**| 43.6 | **60.2** |
|
| 104 |
+
|
| 105 |
+
---
|
| 106 |
|
| 107 |
### Explanation of Groups
|
| 108 |
- **Math Tasks**:
|
| 109 |
- Covers **MathHard**, **Math**, and **GSM8K**, measuring the model's performance on mathematical reasoning and problem-solving tasks.
|
| 110 |
- **Original Capabilities**:
|
| 111 |
- Includes **ARC**, **GPQA**, **MMLU**, and **MMLUP**, reflecting the model’s ability to handle general reasoning and knowledge benchmarks.
|
|
|
|
|
|
|
|
|