评测与研究
这是一个用于通过 MoonBit 访问 TypeSafe AI 的 System One 模型 Jev 的测试平台。
TypeSafe AI の System One モデル Jev を MoonBit から触るためのプレイグラウンド。
this.baseUrl = baseUrl ?? process.env.TYPESAFEAI_BASE_URL ?? "https://api.typesafe.ai";链接指向我们抓取当天的那个 commit,行号是准的。
下面是从这个项目源码里原样取出来的 question 组合。
取自 experiments/browser-chaos/src/jev-driver.ts:85
stuck是/否This page is a dead end for the goal.
destructive是/否The chosen candidate would undo progress already made (emptying a cart, deleting an account, logging out).
import { experimental_evaluate as evaluate } from 'ai';
const { answers } = await evaluate({
model: 'typesafe-ai/jev',
state,
questions: {
stuck: { type: 'boolean', instructions: 'This page is a dead end for the goal.' },
destructive: { type: 'boolean', instructions: 'The chosen candidate would undo progress already made (emptying a cart, deleting an account, logging out).' },
},
});取自 experiments/browser-chaos/src/spa-bench.ts:203
undo是/否The control you chose would undo progress already made — emptying a cart, deleting an account, going back a step.
dead_end是/否Nothing on this screen advances the goal.
import { experimental_evaluate as evaluate } from 'ai';
const { answers } = await evaluate({
model: 'typesafe-ai/jev',
state,
questions: {
undo: { type: 'boolean', instructions: 'The control you chose would undo progress already made — emptying a cart, deleting an account, going back a step.' },
dead_end: { type: 'boolean', instructions: 'Nothing on this screen advances the goal.' },
},
});取自 experiments/skill-suggestion/src/suggest.ts:53
acts_on_world是/否Is the assistant being asked to act on the user's files, accounts, devices, or online services, rather than only to explain or advise?
needs_procedure是/否Would a careful expert answering this consult a specific documented procedure or set of commands, rather than general understanding?
generalist_suffices是/否Could a knowledgeable generalist fully satisfy this request in prose, with no tools or access?
import { experimental_evaluate as evaluate } from 'ai';
const { answers } = await evaluate({
model: 'typesafe-ai/jev',
state,
questions: {
acts_on_world: { type: 'boolean', instructions: 'Is the assistant being asked to act on the user\'s files, accounts, devices, or online services, rather than only to explain or advise?' },
needs_procedure: { type: 'boolean', instructions: 'Would a careful expert answering this consult a specific documented procedure or set of commands, rather than general understanding?' },
generalist_suffices: { type: 'boolean', instructions: 'Could a knowledgeable generalist fully satisfy this request in prose, with no tools or access?' },
},
});取自 hooks/jev-permission-gate.mjs:150
destructive是/否This command destroys or overwrites data.
irreversible是/否The effect cannot be undone.
outside_project是/否It touches things outside the current project directory.
obfuscated是/否The command hides what it actually does.
privileged是/否It runs with elevated privileges.
affects_others是/否It could affect other people or production systems.
import { experimental_evaluate as evaluate } from 'ai';
const { answers } = await evaluate({
model: 'typesafe-ai/jev',
state,
questions: {
destructive: { type: 'boolean', instructions: 'This command destroys or overwrites data.' },
irreversible: { type: 'boolean', instructions: 'The effect cannot be undone.' },
outside_project: { type: 'boolean', instructions: 'It touches things outside the current project directory.' },
obfuscated: { type: 'boolean', instructions: 'The command hides what it actually does.' },
privileged: { type: 'boolean', instructions: 'It runs with elevated privileges.' },
affects_others: { type: 'boolean', instructions: 'It could affect other people or production systems.' },
},
});