Verbatim Spans: Multi-Domain Training Data for Query-Conditioned Evidence Selection

Verbatim Spans is a multi-domain training dataset for query-conditioned extractive evidence selection. The dataset combines three sources covering distinct domains and annotation conventions, including a silver-standard subset of 20,916 training and 2,319 validation rows from NLP research papers. It was created by KRLabsOrg and last updated on Hugging Face in June 2026.

Use Cases

• Training models for extractive evidence selection based on the described query-passage-answer task.
• Benchmarking the performance of large language models on multi-domain evidence extraction based on the three distinct source domains.
• Developing and evaluating retrieval-augmented generation (RAG) pipelines based on the dataset's focus on verbatim substring selection.
• Studying annotation conventions and domain transfer in NLP tasks based on the combined sources with distinct conventions.

Strengths

• Includes a silver-standard subset of 20,916 training and 2,319 validation rows.
• Combines three distinct source domains, which likely provides diversity in content and annotation style.
• Task is clearly defined as selecting verbatim substrings from a passage that support an answer to a given question.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the full combined dataset is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality and the specifics of the other two sources require manual inspection after download.

Provenance

Source

Combines three sources, including a silver-standard subset from NLP research papers generated by KRLabsOrg using Qwen 3.6 35B.

Collection Method

Likely created through automated annotation (e.g., using a large language model) and combination of existing benchmarks.

Freshness

Last updated 2026-06-02 13:25:01; freshness should be verified.