Query Circuit Dataset: Preprocessed Score Matrices for Language Model Interpretability

Preprocessed intermediate data reproduces results from the ICML 2026 paper 'Query Circuits: Explaining How Language Models Answer User Prompts'. The dataset contains score matrices for queries, including the IOI dataset and others with generated paraphrases. It was uploaded by author tony10101105 and last updated on May 29, 2026.

Use Cases

• Reproducing interpretability experiments based on the preprocessed score matrices.
• Analyzing model behavior on specific queries using the provided query IDs.
• Studying the effect of query paraphrases on model activation patterns.
• Benchmarking new interpretability methods against the established IOI and ARC Challenge tasks.

Strengths

• Data is directly linked to a peer-reviewed ICML 2026 publication, providing a clear research context.
• Includes multiple established interpretability benchmarks like IOI and ARC Challenge.
• Files are organized by query ID and paraphrase ID, suggesting a structured format.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and overall dataset scale are unknown, which may limit suitability assessment.
• The description references a full description on an external page, requiring a click-through for complete details.

Provenance

Source

tony10101105

Collection Method

Created as preprocessed intermediate data for reproducing academic paper results.

Freshness

Last updated 2026-05-29 07:46:33