LongDS-Bench is a benchmark for evaluating long-horizon, multi-turn agentic data analysis. It contains 68 tasks constructed from real-world Kaggle notebooks and datasets, spanning 2,225 turns. The dataset was created by zjunlp and was last updated on 2026-05-31.
Use Cases
- Benchmarking agent performance on long-horizon data analysis tasks based on the 68 constructed tasks.
- Testing an agent's ability to maintain evolving analytical states based on the description of filters, metric definitions, and intermediate tables.
- Evaluating multi-turn interaction consistency based on the dataset's focus on tasks spanning 2,225 turns.
Strengths
- Contains 68 tasks constructed from real-world Kaggle notebooks and datasets.
- Spans 2,225 turns, providing a substantial testbed for multi-turn interactions.
Limitations
- Column-level documentation is absent; field semantics must be inferred after download.
- Row count is unknown, which may limit suitability assessment.
Provenance
- Source
- zjunlp
- Collection Method
- Constructed from real-world Kaggle notebooks and datasets.
- Freshness
- Last updated 2026-05-31 07:06:37.