sheryc/DROP_care

Dataset Card for DROP for CARE training

     

Dataset Description

This dataset is a processed version of DROP (Discrete Reasoning Over Paragraphs) adapted for training language models with native retrieval-augmented reasoning capabilities in the CARE framework. It serves as the "easy" dataset (D_easy) in the curriculum learning strategy for reinforcement learning training.

Key Features

• 77,409 training instances and 9,536 validation instances
• Formatted for reasoning-based QA with context and prompts
• Discrete reasoning tasks including counting, sorting, and arithmetic operations
• Structured for curriculum learning as the initial training phase before MS MARCO

Dataset Structure

Data Fields

• section_id: Unique identifier for the passage section
• query_id: Unique identifier for the question
• passage: Original paragraph containing information needed to answer
• question: Question requiring discrete reasoning over the passage
• answers_spans: Dictionary containing:
  • spans: List of answer text spans
  • types: List of answer types (e.g., number, date, span)
• context: Formatted context for model input
• prompt: Complete prompt combining context and question
• answer: Processed answer string
• answer_type: Type classification of the answer

Data Format Example

{
  "passage": "Hoping to rebound from their loss to the Patriots, the Raiders stayed at home for a Week 16 duel with the Houston Texans...",
  "question": "Who scored the first touchdown of the game?",
  "answer": "Chaz Schilens",
  "answer_type": "span",
  "context": "Context: Hoping to rebound from their loss to the Patriots...",
  "prompt": "Context: [passage]\n\nQuestion: Who scored the first touchdown of the game?"
}

Data Splits

Split	Number of Examples
Train	77,409
Validation	9,536

Usage

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("sheryc/DROP_care")
train_data = dataset["train"]
val_data = dataset["validation"]

Training Integration

# Example usage in curriculum learning
class CurriculumTrainer:
    def __init__(self):
        self.drop_data = load_dataset("sheryc/DROP_care")["train"]
        self.msmarco_data = load_dataset("sheryc/MSMARCO_care")["train"]
        
    def get_batch(self, step, total_steps):
        # Linear curriculum: start with DROP, gradually add MS MARCO
        alpha = max(0, 1 - step / total_steps)
        if random.random() < alpha:
            return self.sample_from_drop()
        else:
            return self.sample_from_msmarco()

License

Please refer to the original DROP dataset license.

Citation

If you use this dataset, please cite both the original DROP 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}
}

@inproceedings{Dua2019DROP,
  author={Dheeru Dua and Yizhong Wang and Pradeep Dasigi and Gabriel Stanovsky and Sameer Singh and Matt Gardner},
  title={DROP: A Reading Comprehension Benchmark Requiring Discrete Reasoning Over Paragraphs},
  booktitle={Proc. of NAACL},
  year={2019}
}

Contact

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