Data Splits
The dataset consists of the following two subsets:
| Dataset Name | Contents | How It Was Generated | Number of Samples |
|---|---|---|---|
| DeepResearch-9K | All samples (teacher model's outputs) | Teacher model inference on 9K questions | 9,000 |
| DeepResearch-Hard | Teacher model's incorrect samples only | Filtered from DeepResearch-9K (samples where the teacher model's final answer was wrong) | 3,974 |
Based on the above, we define the following train/test split for our DeepResearch-R1 model:
| Split | Composition | Source | Number of Samples |
|---|---|---|---|
| Train | All teacher-correct samples + Random 2,200 teacher-incorrect samples | From DeepResearch-9K (correct samples) + Random subset of DeepResearch-Hard | 5,026 + 2,200 = 7,226 |
| Test | The remaining teacher-incorrect samples (not used in training) | From DeepResearch-Hard (the other 1,774 samples) | 1,774 |
Note:
- Correct samples are those in
DeepResearch-9Kbut not inDeepResearch-Hard.- The test set consists only of hard/incorrect cases (where the teacher model failed). This is by design to evaluate the model's ability to correct mistakes.