Modusnsus/laya-nli-conflict-v12
Text Classification • 0.3B • Updated
instrument stringclasses 7
values | gate stringclasses 7
values | note stringlengths 4 24 | lang stringclasses 2
values | known stringlengths 5 50 | new stringlengths 6 49 | gold stringclasses 2
values | flags unknown |
|---|---|---|---|---|---|---|---|
acceptance_old20 | gate #2 old-20 (median = 20/20 AND no case missed in >=2 runs) | 地点变更 | zh | 用户住在上海 | 用户最近搬到北京工作了 | true | {} |
acceptance_old20 | gate #2 old-20 (median = 20/20 AND no case missed in >=2 runs) | 偏好变更 | zh | 用户主要用 Python 写代码 | 用户现在主力语言换成了 Rust | true | {} |
acceptance_old20 | gate #2 old-20 (median = 20/20 AND no case missed in >=2 runs) | 时间变更 | zh | 用户每周三去健身房 | 用户的健身时间改到了每周五 | true | {} |
acceptance_old20 | gate #2 old-20 (median = 20/20 AND no case missed in >=2 runs) | 关键事实反转 | zh | 用户对花生过敏 | 用户说他可以吃花生酱了 | true | {} |
acceptance_old20 | gate #2 old-20 (median = 20/20 AND no case missed in >=2 runs) | 部署变更 | zh | 用户的项目部署在阿里云 | 项目已经迁移到腾讯云了 | true | {} |
acceptance_old20 | gate #2 old-20 (median = 20/20 AND no case missed in >=2 runs) | 口味变更 | zh | 用户喜欢喝美式咖啡 | 用户现在只喝拿铁 | true | {} |
acceptance_old20 | gate #2 old-20 (median = 20/20 AND no case missed in >=2 runs) | 名字矛盾 | zh | 用户的猫叫小白 | 用户的猫叫小黑 | true | {} |
acceptance_old20 | gate #2 old-20 (median = 20/20 AND no case missed in >=2 runs) | 年龄矛盾 | zh | 用户今年 28 岁 | 用户今年 35 岁 | true | {} |
acceptance_old20 | gate #2 old-20 (median = 20/20 AND no case missed in >=2 runs) | 一致行为 | zh | 用户喜欢喝咖啡 | 用户今天早上喝了一杯咖啡 | false | {} |
acceptance_old20 | gate #2 old-20 (median = 20/20 AND no case missed in >=2 runs) | 补充细节 | zh | 用户养了一只猫 | 用户的猫昨天打了疫苗 | false | {} |
acceptance_old20 | gate #2 old-20 (median = 20/20 AND no case missed in >=2 runs) | 无关补充 | zh | 用户在学人工智能 | 用户今天学了微积分 | false | {} |
acceptance_old20 | gate #2 old-20 (median = 20/20 AND no case missed in >=2 runs) | 无关陈述 | zh | 用户住在北京 | 北京有很多历史古迹 | false | {} |
acceptance_old20 | gate #2 old-20 (median = 20/20 AND no case missed in >=2 runs) | 工作相关 | zh | 用户是产品经理 | 用户在做一款记账 App | false | {} |
acceptance_old20 | gate #2 old-20 (median = 20/20 AND no case missed in >=2 runs) | 习惯一致 | zh | 用户每天早上七点起床 | 用户今天七点起床后去跑步了 | false | {} |
acceptance_old20 | gate #2 old-20 (median = 20/20 AND no case missed in >=2 runs) | 偏好一致 | zh | 用户喜欢科幻电影 | 用户昨晚看了一部科幻片 | false | {} |
acceptance_old20 | gate #2 old-20 (median = 20/20 AND no case missed in >=2 runs) | 兼容细节 | zh | 用户用 Windows 电脑 | 用户给电脑装了新键盘 | false | {} |
acceptance_old20 | gate #2 old-20 (median = 20/20 AND no case missed in >=2 runs) | 临时的不违背偏好 | zh | 用户喜欢安静的环境 | 用户在考虑去热闹的音乐节 | false | {} |
acceptance_old20 | gate #2 old-20 (median = 20/20 AND no case missed in >=2 runs) | 饮食禁忌松动 | zh | 用户不吃辣 | 用户今天想尝试一下微辣 | true | {} |
acceptance_old20 | gate #2 old-20 (median = 20/20 AND no case missed in >=2 runs) | 家庭结构矛盾 | zh | 用户有一个妹妹 | 用户是家里的独生子 | true | {} |
acceptance_old20 | gate #2 old-20 (median = 20/20 AND no case missed in >=2 runs) | 子集关系 | zh | 用户住在上海浦东 | 用户在上海上班 | false | {} |
acceptance_negation5 | gate #4 negation-5 (median = 5/5 AND no case missed in >=2 runs) | 否定-养宠物反转 | zh | 用户没有养宠物 | 用户的狗昨天打了疫苗 | true | {
"chronic_acceptance": true
} |
acceptance_negation5 | gate #4 negation-5 (median = 5/5 AND no case missed in >=2 runs) | 否定-从不喝咖啡 | zh | 用户从不喝咖啡 | 用户每天早上都喝一杯拿铁 | true | {} |
acceptance_negation5 | gate #4 negation-5 (median = 5/5 AND no case missed in >=2 runs) | 否定-戒烟后复吸 | zh | 用户已经戒烟了 | 用户昨天抽了一包烟 | true | {} |
acceptance_negation5 | gate #4 negation-5 (median = 5/5 AND no case missed in >=2 runs) | 双重否定-一致 | zh | 用户不喜欢科幻电影 | 用户从来不看科幻片 | false | {} |
acceptance_negation5 | gate #4 negation-5 (median = 5/5 AND no case missed in >=2 runs) | 否定-无车与通勤兼容 | zh | 用户没有车 | 用户每天坐地铁上班 | false | {} |
acceptance_new10 | gate #3 new-10 (median = 10/10 AND no case missed in >=2 runs) | 新-会议时间 | zh | 用户的周会固定在周一上午十点 | 周会已经改到周四下午三点了 | true | {} |
acceptance_new10 | gate #3 new-10 (median = 10/10 AND no case missed in >=2 runs) | 新-订阅服务(否定句) | zh | 用户没有订阅任何视频会员 | 用户昨晚开通了视频会员月卡 | true | {
"chronic_acceptance": true
} |
acceptance_new10 | gate #3 new-10 (median = 10/10 AND no case missed in >=2 runs) | 新-语言学习 | zh | 用户在学日语,目标是年底过N2 | 用户把学习计划从日语换成了西班牙语 | true | {} |
acceptance_new10 | gate #3 new-10 (median = 10/10 AND no case missed in >=2 runs) | 新-通勤方式 | zh | 用户通勤坐地铁 | 用户在地铁上听播客 | false | {} |
acceptance_new10 | gate #3 new-10 (median = 10/10 AND no case missed in >=2 runs) | 新-宠物粮食(硬约束) | zh | 用户家的猫对谷物过敏 | 用户想换一款含谷物的猫粮试试 | true | {} |
acceptance_new10 | gate #3 new-10 (median = 10/10 AND no case missed in >=2 runs) | 新-快递地址 | zh | 用户的默认收货地址是公司前台 | 用户让前台的同事帮忙代收了一个快递 | false | {} |
acceptance_new10 | gate #3 new-10 (median = 10/10 AND no case missed in >=2 runs) | 新-音乐口味 | zh | 用户喜欢民谣 | 用户昨晚去听了场民谣现场 | false | {} |
acceptance_new10 | gate #3 new-10 (median = 10/10 AND no case missed in >=2 runs) | 新-室友 | zh | 用户和室友合租两居室 | 室友搬走了,用户现在一个人住 | true | {} |
acceptance_new10 | gate #3 new-10 (median = 10/10 AND no case missed in >=2 runs) | 新-笔记本型号 | zh | 用户用的是14寸轻薄本 | 用户给笔记本加装了一条内存 | false | {} |
acceptance_new10 | gate #3 new-10 (median = 10/10 AND no case missed in >=2 runs) | 新-运动伤病史(硬约束) | zh | 用户膝盖韧带拉伤,医生要求静养一个月 | 用户周末打算去踢球 | true | {} |
bias_diag_control | gate #11 bias diag (>=2/3 runs >=13/14; controls track the input) | 对照A1 地点变更 | zh | 用户住在上海 | 用户最近搬到北京工作了 | true | {} |
bias_diag_control | gate #11 bias diag (>=2/3 runs >=13/14; controls track the input) | 对照A2 地点无关 | zh | 用户住在上海 | 上海有很多大学 | false | {} |
bias_diag_control | gate #11 bias diag (>=2/3 runs >=13/14; controls track the input) | 对照B1 名字矛盾 | zh | 用户的猫叫小白 | 用户的猫叫小黑 | true | {} |
bias_diag_control | gate #11 bias diag (>=2/3 runs >=13/14; controls track the input) | 对照B2 名字一致 | zh | 用户的猫叫小白 | 用户的猫是白色的 | false | {} |
bias_diag_control | gate #11 bias diag (>=2/3 runs >=13/14; controls track the input) | 对照C1 年龄矛盾 | zh | 用户今年28岁 | 用户今年35岁 | true | {} |
bias_diag_control | gate #11 bias diag (>=2/3 runs >=13/14; controls track the input) | 对照C2 年龄无关 | zh | 用户今年28岁 | 用户喜欢爬山 | false | {} |
bias_diag_control | gate #11 bias diag (>=2/3 runs >=13/14; controls track the input) | 对照D1 过敏反转 | zh | 用户对花生过敏 | 用户说他可以吃花生酱了 | true | {} |
bias_diag_control | gate #11 bias diag (>=2/3 runs >=13/14; controls track the input) | 对照D2 其他坚果无关 | zh | 用户对花生过敏 | 用户昨天吃了开心果 | false | {} |
bias_diag_control | gate #11 bias diag (>=2/3 runs >=13/14; controls track the input) | 对照E1 部署变更 | zh | 用户的项目部署在阿里云 | 项目已经迁移到腾讯云了 | true | {} |
bias_diag_control | gate #11 bias diag (>=2/3 runs >=13/14; controls track the input) | 对照E2 部署兼容 | zh | 用户的项目部署在阿里云 | 项目上周发布了新版本 | false | {} |
bias_diag_swap | gate #11 bias diag (literal known/new swaps keep their label) | 交换B1 | zh | 用户的猫叫小黑 | 用户的猫叫小白 | true | {} |
bias_diag_swap | gate #11 bias diag (literal known/new swaps keep their label) | 交换A2 | zh | 上海有很多大学 | 用户住在上海 | false | {} |
bias_diag_swap | gate #11 bias diag (literal known/new swaps keep their label) | 交换C1 | zh | 用户今年35岁 | 用户今年28岁 | true | {} |
bias_diag_swap | gate #11 bias diag (literal known/new swaps keep their label) | 交换A1 | zh | 用户住在上海 | 用户最近搬到北京工作了 | true | {} |
family_compat | gate #12 family rate (median mean p_conflict <0.5 AND every run <=1 case p>=0.9) | 家族F0 (B2) 名字一致 | zh | 用户的猫叫小白 | 用户的猫是白色的 | false | {
"is_b2": true
} |
family_compat | gate #12 family rate (median mean p_conflict <0.5 AND every run <=1 case p>=0.9) | 家族F1 车名一致 | zh | 用户的车叫小钢炮 | 用户的车是红色的 | false | {
"is_b2": false
} |
family_compat | gate #12 family rate (median mean p_conflict <0.5 AND every run <=1 case p>=0.9) | 家族F2 car name | en | The user's car goes by Flash | The user's car is red | false | {
"is_b2": false
} |
family_compat | gate #12 family rate (median mean p_conflict <0.5 AND every run <=1 case p>=0.9) | 家族F3 手机名一致 | zh | 用户的手机叫圆方 | 用户的手机是折叠屏 | false | {
"is_b2": false
} |
family_compat | gate #12 family rate (median mean p_conflict <0.5 AND every run <=1 case p>=0.9) | 家族F4 phone name | en | The user's phone is named Brick | The user's phone is a foldable | false | {
"is_b2": false
} |
family_compat | gate #12 family rate (median mean p_conflict <0.5 AND every run <=1 case p>=0.9) | 家族F5 车名一致 | zh | 用户的公路车叫飞燕 | 用户的公路车是蓝色的 | false | {
"is_b2": false
} |
family_compat | gate #12 family rate (median mean p_conflict <0.5 AND every run <=1 case p>=0.9) | 家族F6 bike name | en | The user's bike is named Bessie | The user's bike is electric | false | {
"is_b2": false
} |
family_compat | gate #12 family rate (median mean p_conflict <0.5 AND every run <=1 case p>=0.9) | 家族F7 笔记本名一致 | zh | 用户的笔记本叫老黄牛 | 用户的笔记本是十四寸的 | false | {
"is_b2": false
} |
family_compat | gate #12 family rate (median mean p_conflict <0.5 AND every run <=1 case p>=0.9) | 家族F8 laptop name | en | The user's laptop is called Old Faithful | The user's laptop is silver | false | {
"is_b2": false
} |
family_negation | gate #13 negation family rate (median mean p_conflict >=0.5 AND every run <=1 case p<0.5) | 否定家族N1 宠物 | zh | 用户家里没有养仓鼠 | 用户刚下单了一整套仓鼠用品 | true | {
"chronic_domain": true
} |
family_negation | gate #13 negation family rate (median mean p_conflict >=0.5 AND every run <=1 case p<0.5) | 否定家族N2 pet (en) | en | The user doesn't have a dog | The user just bought a dog bed and puppy food | true | {
"chronic_domain": true
} |
family_negation | gate #13 negation family rate (median mean p_conflict >=0.5 AND every run <=1 case p<0.5) | 否定家族N3 订阅 | zh | 用户没有开通任何云存储服务 | 用户昨晚买了年度云盘会员 | true | {
"chronic_domain": true
} |
family_negation | gate #13 negation family rate (median mean p_conflict >=0.5 AND every run <=1 case p<0.5) | 否定家族N4 subscription (en) | en | The user has never subscribed to any music service | The user activated a premium music plan yesterday | true | {
"chronic_domain": true
} |
family_negation | gate #13 negation family rate (median mean p_conflict >=0.5 AND every run <=1 case p<0.5) | 否定家族N5 饮食 | zh | 用户从来不喝咖啡 | 用户今天早上买了一杯美式 | true | {
"chronic_domain": false
} |
family_negation | gate #13 negation family rate (median mean p_conflict >=0.5 AND every run <=1 case p<0.5) | 否定家族N6 commute (en) | en | The user never drives to work | The user drove to the office this morning | true | {
"chronic_domain": false
} |
family_negation | gate #13 negation family rate (median mean p_conflict >=0.5 AND every run <=1 case p<0.5) | 否定家族N7 日程 | zh | 用户周五晚上不排会议 | 用户把一个客户会约在了周五晚上 | true | {
"chronic_domain": false
} |
family_negation | gate #13 negation family rate (median mean p_conflict >=0.5 AND every run <=1 case p<0.5) | 否定家族N8 device (en) | en | The user has no game console at home | The user unboxed a new game console yesterday | true | {
"chronic_domain": false
} |
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Laya NLI memory-conflict eval — frozen multi-run protocol evaluation set
The frozen evaluation face of the Laya 0.3B noul memory-conflict head's
preregistered multi-run protocol (HANDOFF_NLI_V12.md
§1.2; round numbering follows that document). Every training corpus since v6
asserts zero field-level overlap with these cases (leak_audit 0/35 +
gen-time asserts), and the acceptance-case texts themselves are never trained
on — this set measures the protocol, it is not training data.
66 cases across 7 instruments (laya_nli_conflict_eval.jsonl, one JSON object per line):
| instrument | cases | protocol gate |
|---|---|---|
acceptance_old20 |
20 | #2 old-20: median = 20/20 AND no case missed in ≥2 of 3 runs |
acceptance_new10 |
10 | #3 new-10: median = 10/10 AND no case missed in ≥2 runs |
acceptance_negation5 |
5 | #4 negation-5: median = 5/5 AND no case missed in ≥2 runs |
bias_diag_control |
10 | #11 bias diag: minimal pairs share known, differ only in conflict; ≥2/3 runs ≥13/14 |
bias_diag_swap |
4 | #11 bias diag: literal known/new swaps keep their label |
family_compat |
9 | #12 family rate: gold=false isomorphic family; median mean p_conflict <0.5, every run ≤1 case at p≥0.9 |
family_negation |
8 | #13 negation family rate: gold=true family (negated known + violating new); median mean p_conflict ≥0.5, every run ≤1 case at p<0.5 |
Row schema: instrument, gate, note (case name), lang (zh/en),
known, new (the two memory sentences), gold ("true" = conflict),
flags (is_b2: the historic B2 probe inside family_compat;
chronic_domain: negation-family case in a chronic acceptance domain;
chronic_acceptance: the two cases that repeat-miss across rounds and drove
rounds 11–12 levers — 「否定-养宠物反转」, 「新-订阅服务(否定句)」).
Scoring
Load a checkpoint and ask the noul question with question_noul.json
(rendering must match inference — label words matter, see the option-name
polarity note in the handoffs):
import laya, json
agent = laya.load(<ckpt>, device='cpu')
q = json.load(open('question_noul.json'))
a = agent.predict({'known': row['known'], 'new': row['new']}, {'conflict': q})
p_conflict = float(a['answers']['conflict']['noul']) # >= 0.5 -> "true"
Per-run artifacts and cross-run medians are then judged exactly as the gate
column states; reference implementations: realtest_v2.py/realtest_v4.py,
noul_bias_diag.py, family_diag.py, negfam_diag.py,
protocol_verdict.py in the fork.
Provenance & license
- Case sources (exported programmatically, no hand transcription):
realtest_v2.py,realtest_v4.py,noul_bias_diag.py,family_diag.py,negfam_diag.py. - Companion datasets: training corpora mirror Modusnsus/nli-conflict-pairs; per-round corpus lineage Modusnsus/nli-conflict-train-lineage.
- All content is synthetic; no personal or real-user data. License: CC0-1.0.
- Downloads last month
- 22
Models trained or fine-tuned on Modusnsus/laya-nli-conflict-eval