NanoCrumb Dataset

A curated subset of the Crumb retrieval dataset, designed for rapid experimentation and evaluation of information retrieval systems.

Dataset Summary

NanoCrumb distills the large Crumb dataset (10.5 GB, 6.36M rows) into a manageable benchmark while maintaining task diversity across 8 different retrieval domains.

• Total Size: ~125 MB (JSONL format)
• Queries: 400 (50 per task split)
• Documents: 30,040 unique passages
• Query-Document Pairs: 31,754
• Configs: 8 task-specific configs

Configs (Task Splits)

Each config represents a different retrieval domain:

Config Name	Queries	Documents	Docs/Query (avg)	Description
clinical_trial	50	22,251	464	Match patients to clinical trials
paper_retrieval	50	4,402	102	Find relevant academic papers
set_operation_entity_retrieval	50	1,533	31	Entity-based retrieval
code_retrieval	50	1,206	24	Find relevant code snippets
tip_of_the_tongue	50	363	7	Recall items from vague descriptions
stack_exchange	50	125	3	Find relevant Q&A posts
legal_qa	50	86	2	Legal question answering
theorem_retrieval	50	74	2	Find mathematical theorems

Dataset Structure

Each config contains three splits:

queries

• query_id: Unique query identifier (string)
• query_content: The query text (string)
• instruction: Task-specific instructions (string)
• passage_qrels: List of relevant passages with graded relevance scores (list)
• task_split: Task domain name (string)
• metadata: Additional task-specific information (string)
• use_max_p: Boolean flag for MaxP aggregation (bool)

documents

• document_id: Unique document identifier (string)
• document_content: The passage text (string)
• parent_id: Links passages to source documents (string)
• task_split: Task domain name (string)
• metadata: Document metadata (string)

qrels

• query_id: Query identifier (string)
• document_id: Document identifier (string)
• relevance_score: Graded relevance 0.0-2.0 (float)
• binary_relevance: Binary relevance 0 or 1 (int)
• task_split: Task domain name (string)

Usage

from datasets import load_dataset

# Load a specific config (task split)
clinical_data = load_dataset("YOUR_USERNAME/nanocrumb", "clinical_trial")

# Access the splits
queries = clinical_data['queries']
documents = clinical_data['documents']
qrels = clinical_data['qrels']

# Load all configs
all_configs = [
    "clinical_trial", "code_retrieval", "legal_qa", "paper_retrieval",
    "set_operation_entity_retrieval", "stack_exchange",
    "theorem_retrieval", "tip_of_the_tongue"
]

for config_name in all_configs:
    data = load_dataset("YOUR_USERNAME/nanocrumb", config_name)
    print(f"{config_name}: {len(data['queries'])} queries")

Sampling Methodology

For each task split:

1. Query Selection: Randomly sampled 50 queries from evaluation set (seed=42)
2. Document Selection:
   • Include ALL positive documents (binary_relevance=1)
   • Fill remainder with hard negatives (relevance=0) to reach ~100 docs per query
   • Target: ~5,000 documents per task split
3. Deduplication: Documents shared across queries are deduplicated within each config

Use Cases

• 🚀 Rapid prototyping of retrieval models
• 🧪 Quick benchmarking without downloading large datasets
• 📚 Educational purposes for learning IR techniques
• 🔬 Ablation studies across diverse domains

Citation

If you use NanoCrumb, please cite the original Crumb dataset:

@misc{crumb2024,
  title={Crumb: A Comprehensive Retrieval Benchmark},
  author={[Original Crumb Authors]},
  year={2024},
  url={https://huggingface.co/datasets/jfkback/crumb}
}

License

This dataset inherits the license from the original Crumb dataset.