name: "ACI-Bench"
description: >-
  Attention Clarity Index Bench: measures whether small language models
  (1M-150M parameters) allocate attention*gradient saliency toward
  causally-relevant sentences over same-domain distractors, via a
  white-box internals probe rather than answer-correctness grading.
tasks:
  - id: "aci_bench"
    config: "default"
    split: "test"
    metric: "aciScore"
    reference_implementation: "https://huggingface.co/datasets/AxiomicLabs/ACI-Bench/blob/main/evaluation_harness.py"