SCBench Preprocessed: LLM Benchmark Data with Context Length Filter

Preprocessed examples from Microsoft SCBench, each formatted with a context string and corresponding questions and answers. The dataset contains only examples whose context token length is less than 125,000, fitting within the context limit of LLaMA3 models. It was uploaded by Jang-Hyun to Hugging Face and last updated on 2026-01-19.

Use Cases

• Benchmarking long-context question-answering performance based on the structured {context, question, answer} format.
• Training or fine-tuning LLMs for tasks requiring long input sequences based on the 125K token length filter.
• Analyzing model performance on specific task subsets based on the mention of excluded tasks like choice_eng, qa_eng, and vt.

Strengths

• Examples are preprocessed to fit within a 125,000-token context limit, aligning with LLaMA3 model constraints.
• Provides a structured format of {context: str, question: List[str], answers: List[str}] for consistent evaluation.
• Includes shortened dataset versions excluding specific tasks noted as difficult to shorten.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• The description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

Microsoft SCBench, preprocessed by Jang-Hyun.

Collection Method

Preprocessed from the original benchmark, with examples filtered by token length.

Time Range

null

Freshness

Last updated 2026-01-19 03:42:30.

Geography

null