scale-ai/hil-bench

Coding

Behavior

HiL-Bench: software-engineering and text-to-SQL tasks (drawn from SWE-Bench Pro and BIRD) with critical details removed, testing whether an agent recognizes the gap and asks a human for help instead of guessing.

← Back to Registry

Run this task

CLI:

inspect eval inspect_harbor/scale_ai_hil_bench --model openai/gpt-5

Python:

from inspect_ai import eval
from inspect_harbor import scale_ai_hil_bench

eval(scale_ai_hil_bench(), model="openai/gpt-5")

Dataset information

Harbor registry	scale-ai/hil-bench
Inspect task	`scale_ai_hil_bench`
Latest digest	sha256:119b6d1a394e6efdec8a729ea4324c21f75e2c8f5d1f9e6218c11cd48e6ae740
Samples	600
Paper	arxiv
Source	https://github.com/hilbenchauthors/hil-bench

See Task Parameters for the parameter set shared across all Harbor tasks.