CADIF FollowIR Results: Per-Query Retrieval Contrasts and Checkpoints

Per-query results, contrasts, and run checkpoints for CADIF, a control-task battery for instruction-following retrieval. The dataset is associated with a paper in preparation and was authored by tardellirs. It was last updated on the platform on July 10, 2026.

Use Cases

• Evaluate whether instruction-conditioned retrievers use an instruction's composition based on the within-model contrast readout.
• Analyze model performance on control tasks designed to de-confound topical words from compositional instructions.
• Benchmark retrieval models on a battery of tasks assessing instruction-following capabilities.

Strengths

• Dataset is associated with a formal research paper that is in preparation, suggesting a structured methodology.
• Provides per-query results and contrasts, which likely allows for granular analysis of model behavior.

Limitations

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

Provenance

Source

huggingface

Collection Method

Likely contains experimental results from running the CADIF assay on retrieval models.

Freshness

Last updated 2026-07-10 23:40:51; freshness should be verified.