dingchenhui0618-arch/dsh-turing-table ↗★ 0
dsh-turing-table
Turing table for DeepSeek Harness: the seats at one table are played by agents on different model configs, you question them and cross-check their answers to find which seats are AI, and every round exports a reproducible per-seat JSONL transcript plus a verdict record.
安装
$
npx -p @deepseek-ai/dsh dsh plugin --profile web add github:dingchenhui0618-arch/dsh-turing-table说明文档
阅读完整 README ↗配置
{
"seed": "my-round-0001", // 不给则用时间戳生成,并显示在面板上(对局仍会被完整记录)
"seatCount": 6,
"humanSeats": 1, // 至少 1,且至少留一个 AI 座位
"maxRounds": 8,
"exportDir": "",
"models": [
{ "id": "a", "provider": "deepseek", "model": "", "temperature": 0.3,
"system": "你话少、只说结论。", "style": "书面", "voice": "dry",
"disguise": { "level": "low" } },
{ "id": "b", "provider": "deepseek", "model": "", "temperature": 0.9,
"system": "你爱闲聊。", "style": "口语化、话多", "voice": "warm",
"disguise": { "level": "high", "pressure": "hold" } }
]
}
models 是座位模型池,不是座位表;谁拿到哪个由 seed 决定。
单人默认配置的坦白话:humanSeats: 1 时,除你之外所有座位都是 AI,于是“谁是 AI”在已知信息下等于全选。要让这一局真的有猜的成分,需要 ≥2 个人类座位——多人同桌(同一张桌上几个人各自输入),或者由外部驱动人类座位(host 的 POST /say,或脚本里的 say 步骤)。examples/sample-spec.json 就是一个 6 座 2 人的样例:4 个 AI、1 个人类座位(脚本喂台词)、外加你自己。