sheryc/ms_marco_care

Dataset Card for MS MARCO for CARE Training

     

Dataset Description

This dataset is a processed version of MS MARCO adapted for training language models with native retrieval-augmented reasoning capabilities in the CARE framework. It serves as the "hard" dataset (D_hard) in the curriculum learning strategy for reinforcement learning training.

Key Features

• 808,731 training instances, 101,093 validation instances, and 101,092 test instances
• Real web queries from Bing search with human-generated answers
• Longer contexts and answers compared to DROP dataset
• Complex multi-hop reasoning required for comprehensive answers

Dataset Structure

Data Fields

• query_id: Unique identifier for the query (int32)
• query_type: Classification of query type
• question: Original user query from Bing search
• context: Concatenated passages containing relevant information
• prompt: Complete prompt combining context and question
• answer: Human-generated answer text

Data Format Example

{
  "query_id": 1102432,
  "query_type": "description",
  "question": "what causes blood pressure to rise?",
  "context": "Context: High blood pressure can be caused by several factors including...",
  "prompt": "Context: [passages]\n\nQuestion: what causes blood pressure to rise?",
  "answer": "Blood pressure rises due to factors such as stress, high sodium intake, lack of physical activity..."
}

Data Splits

Split	Number of Examples
Train	808,731
Validation	101,093
Test	101,092

Usage

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("sheryc/ms_marco_care")
train_data = dataset["train"]
val_data = dataset["validation"]
test_data = dataset["test"]

Curriculum Learning Integration

# Example usage as Dhard in curriculum learning
class CurriculumScheduler:
    def __init__(self, total_steps):
        self.drop_data = load_dataset("sheryc/DROP_care")["train"]
        self.msmarco_data = load_dataset("sheryc/ms_marco_care")["train"]
        self.total_steps = total_steps
        
    def get_training_batch(self, current_step, batch_size):
        # Gradually transition from DROP to MS MARCO
        alpha = max(0, 1 - current_step / self.total_steps)
        
        if random.random() < alpha:
            # Sample from DROP (easy)
            return self.sample_batch(self.drop_data, batch_size)
        else:
            # Sample from MS MARCO (hard)
            return self.sample_batch(self.msmarco_data, batch_size)

License

Please refer to the original MS MARCO dataset license and terms of use.

Citation

If you use this dataset, please cite both the original MS MARCO paper and the CARE paper:

@inproceedings{wang2025care,
  title={Improving Context Fidelity via Native Retrieval-Augmented Reasoning},
  author={Wang, Suyuchen and Wang, Jinlin and Wang, Xinyu and Li, Shiqi and Tang, Xiangru and Hong, Sirui and Chang, Xiao-Wen and Wu, Chenglin and Liu, Bang},
  booktitle={Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing},
  year={2025}
}

@article{DBLP:journals/corr/NguyenRSGTMD16,
  author    = {Tri Nguyen and Mir Rosenberg and Xia Song and Jianfeng Gao and 
               Saurabh Tiwary and Rangan Majumder and Li Deng},
  title     = {{MS} {MARCO}: {A} Human Generated MAchine Reading COmprehension Dataset},
  journal   = {CoRR},
  volume    = {abs/1611.09268},
  year      = {2016},
  url       = {http://arxiv.org/abs/1611.09268}
}

Contact

For questions about the dataset or to report issues, please visit the CARE project homepage or contact the authors.