DARS: Scored LLM Generations for Distribution-Aware Routing

AIGNLAI's DARS Routing Benchmark is a collection of scored large language model generations for studying model routing. It provides multiple observations for each query-model pair by varying inputs and decoding processes. The dataset was last updated on June 4, 2026.

Use Cases

• Evaluating model routing algorithms based on multiple scored generations per query.
• Studying the impact of input-side variation (semantically equivalent rewrites) on model performance.
• Analyzing the effect of decoding process variation on generation quality for routing decisions.
• Benchmarking distribution-aware supervision techniques for LLM ensembles.

Strengths

• Provides multiple scored observations per query-model pair, unlike conventional single-label routing datasets.
• Introduces variation through both input-side rewrites and decoding process changes.

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

AIGNLAI via Hugging Face

Collection Method

Likely contains scored LLM generations created for a routing benchmark.

Freshness

Last updated 2026-06-04 07:38:06