Back to datasets
Dataset assetOpen Source CommunityLong Text ProcessingNLP Reasoning Evaluation
RMT-team/babilong-1k-samples
BABILong is a benchmark for evaluating NLP models' ability to handle distributed facts across long documents. It comprises nine configurations corresponding to different sequence lengths (0k, 1k, 2k, 4k, 8k, 16k, 32k, 64k, 128k). The dataset builds on the bAbI fact set and PG‑19 as background text, simulating the task of locating crucial information among extensive irrelevant details. Additionally, it includes ten tasks to assess basic reasoning capabilities.
Source
hugging_face
Created
Nov 28, 2025
Updated
Jun 17, 2024
Signals
110 views
Availability
Linked source ready
Overview
Dataset description and usage context
Dataset Overview
Dataset Name
- BABILong
Configurations
- Configuration Names: 0k, 1k, 2k, 4k, 8k, 16k, 32k, 64k, 128k
- Features:
- question: String
- target: String
- input: String
Size
- Download Size: Varies per configuration, ranging from 8,143,277 bytes to 1,567,936,012 bytes
- Dataset Size: Varies per configuration, ranging from 13,838,997 bytes to 2,532,955,312 bytes
Splits
- Split Names: qa1, qa2, …, qa20
- Example Count: Between 999 and 1,000 per split
- Bytes per Split: Varies with configuration, ranging from 2,801,155 bytes to 507,056,606 bytes
File Paths
- Data files follow the pattern
<config_name>/qa<split_number>-*
Tasks
- Task Types: Include single supporting fact, two supporting facts, three supporting facts, etc.
- Number of Facts: Ranges from 2 to 126 depending on the task
- Supporting Facts: 1 to 3 per task
Intended Use
- Evaluating NLP model performance on handling distributed facts in long documents
Sources
- Uses the bAbI dataset for facts and PG‑19 for background text
License
- Apache 2.0 License, BSD License
Need downstream help?
Pair the dataset with AI analysis and content workflows.
Once the source passes your review, move straight into summarization, transformation, report drafting, or presentation generation with the JuheAI toolchain.