neo4j/

aip-skillbench-24task-sonnet-aipv0_3a3

Raw evaluation-run data.

A head-to-head evaluation of two skill formats on the same tasks:

human-curated — the task's original human-authored skill (prose). aip-from-curated — that same human skill compiled into AIP (a schema-validated execution-graph representation).

The 24-task stratified sample is run as three balanced 8-task cohorts (A, B, C). Each task × mode is run for 5 independent… See the full description on the dataset page: https://huggingface.co/datasets/neo4j/aip-skillbench-24task-sonnet-aipv0_3a3.