userbench/UserBench-train400

Assistants

Coding

UserBench train400 twin: the same 620 held tasks plus 400 leak-safe prior turns per developer. Composer 2.5 multi-label acts score by set Jaccard/IoU.

← Back to Registry

Run this task

CLI:

inspect eval inspect_harbor/userbench_train400 --model openai/gpt-5

Python:

from inspect_ai import eval
from inspect_harbor import userbench_train400

eval(userbench_train400(), model="openai/gpt-5")

Dataset information

Harbor registry	userbench/UserBench-train400
Inspect task	`userbench_train400`
Latest digest	sha256:ed4a2f13efe35bec348b5da764c9ccb290443a9d42d622f055eb701dc5c0d2ab
Samples	620

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