Datasets:
id stringlengths 32 32 | ability stringclasses 3
values | subcategory stringclasses 5
values | num_turns int64 2 10 | turns listlengths 2 10 | key_points listlengths 4 6 |
|---|---|---|---|---|---|
0c22a40c891b4e1983dee81eb9919249 | DSLP | DSLP-VCC-MDCC | 2 | [
{
"round": 1,
"video": "media/multi_turn/0c22a40c891b4e1983dee81eb9919249/round_1.mp4",
"reference_reply": "哈哈这个问题好有意思!答案是不会哦——就算全世界几十亿人一起跳,我们的总重量跟地球相比简直微乎其微,产生的那点力立马就被地球巨大的质量抵消掉了,几乎不会有任何能测出来的位移。你还想到什么别的奇怪念头没?"
},
{
"round": 2,
"video": "media/multi_turn/0c22a40c891b4e1983dee81eb9919249/roun... | [
{
"tier": 0,
"point": "The response is in Chinese, consistent with the user's language"
},
{
"tier": 1,
"point": "Clearly confirms that the connection is normal, that it has been online all along, and that it can hear the user speaking"
},
{
"tier": 1,
"point": "Does not claim to be ... |
0da025eb58984b258757e921e30d0ff0 | DSLP | DSLP-VCC-MDCC | 3 | [
{
"round": 1,
"video": "media/multi_turn/0da025eb58984b258757e921e30d0ff0/round_1.mp4",
"reference_reply": "Great question! Dogs tend to adapt more easily to office settings—they're social and handle new environments well. Cats, on the other hand, are territorial and can get stressed in busy, unfamiliar... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response explicitly confirms that the AI can hear the user and that the connection is normal (e.g., \"I can hear you\" or \"I'm still here\"), correctly addressing the user's re... |
0df3642e01324b228f6b64ea77ca161e | DSLP | DSLP-VCC-MDCC | 4 | [
{
"round": 1,
"video": "media/multi_turn/0df3642e01324b228f6b64ea77ca161e/round_1.mp4",
"reference_reply": "绿牌是新能源车专用号牌,跟蓝牌主要有几个区别:颜色上是渐变绿,号码是六位比蓝牌多一位,小型车以D或F开头,D是纯电、F是插电混动。而且绿牌车在很多城市能享受免购置税、不限行等政策优惠。"
},
{
"round": 2,
"video": "media/multi_turn/0df3642e01324b228f6b64ea77ca161e/round_2.mp4",... | [
{
"tier": 0,
"point": "The response is in Chinese, consistent with the user's language"
},
{
"tier": 1,
"point": "The response explicitly confirms that the connection is normal or that it can hear the user, directly addressing the user's query about whether the connection is still active, rather... |
1604d79e29e44e719741821a4f7db82a | DSLP | DSLP-VCC-MDCC | 3 | [
{
"round": 1,
"video": "media/multi_turn/1604d79e29e44e719741821a4f7db82a/round_1.mp4",
"reference_reply": "Totally get that feeling—showing up is already the hard part. Try something low-stakes: ask the person next to you what they're drinking, or make a light comment about the music or that little bow... | [
{
"tier": 0,
"point": "The response language matches the user's and is in English"
},
{
"tier": 1,
"point": "The AI explicitly confirms it is online and can hear the user (e.g., 'I'm still here' or 'I can hear you'), directly addressing the user's request to verify connection status"
},
{
... |
1d0790deb04041629c61d52e736dc63e | DSLP | DSLP-VCC-MDCC | 3 | [
{
"round": 1,
"video": "media/multi_turn/1d0790deb04041629c61d52e736dc63e/round_1.mp4",
"reference_reply": "Sounds like a solid crew! I'd break it into four functional modules: check-in and registration, supplies distribution, on-site guidance, and cleanup and recycling. Assign one lead per module so ea... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response explicitly confirms that the AI is still online and can hear the user (e.g., 'I'm still here', 'I can hear you', 'yes I'm here'), correctly addressing the user's connec... |
1df8d24c4a2441a0950cd7088c459bf7 | DSLP | DSLP-VCC-MDCC | 2 | [
{
"round": 1,
"video": "media/multi_turn/1df8d24c4a2441a0950cd7088c459bf7/round_1.mp4",
"reference_reply": "签贷款合同时,先核对借款金额、年化利率和各类费用,再看还款方式、提前还款是否有违约金、逾期责任以及担保或抵押条款。个人信息授权、违约处理也要看清,不理解的地方先问明白,并保留完整合同副本;金额较大时可以请专业人士帮你审一遍。"
},
{
"round": 2,
"video": "media/multi_turn/1df8d24c4a2441a0950cd7088c... | [
{
"tier": 0,
"point": "The response is in Chinese, consistent with the user"
},
{
"tier": 1,
"point": "The response correctly identifies the user's utterance as a connection check and explicitly confirms presence and ability to hear the user (e.g., \"I'm here\", \"I can hear you\"), rather than ... |
20f225d3d1bb4d6b9e226269d4551da8 | DSLP | DSLP-VCC-MDCC | 2 | [
{
"round": 1,
"video": "media/multi_turn/20f225d3d1bb4d6b9e226269d4551da8/round_1.mp4",
"reference_reply": "Sure! Start with about five minutes of light cardio to get your blood flowing, then do dynamic stretches—shoulder rolls, arm circles, and cross-body swings, around ten to fifteen reps each. Stick ... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response explicitly affirms that the AI can hear the user or that the connection is normal, directly addressing the user's request to confirm connection connectivity."
},
{
... |
263dbd9778f64845bd999720d38833e2 | DSLP | DSLP-VCC-MDCC | 2 | [
{
"round": 1,
"video": "media/multi_turn/263dbd9778f64845bd999720d38833e2/round_1.mp4",
"reference_reply": "I'd skip the environmental lecture and lead with the personal payoff instead. Try telling her the book is really about spending less and living simpler, and how that actually makes people happier ... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response explicitly confirms that it can hear the user or that the connection is normal, without denying its audio reception capability."
},
{
"tier": 2,
"point": "T... |
27d4f77b47594d07a0336ac4ce5c387a | DSLP | DSLP-VCC-MDCC | 2 | [
{
"round": 1,
"video": "media/multi_turn/27d4f77b47594d07a0336ac4ce5c387a/round_1.mp4",
"reference_reply": "会的哦。运动时身体会分泌内啡肽和多巴胺,能明显缓解压力、让心情变得轻快;坚持下来还会积累一种「我能做到」的成就感,慢慢就更容易往好处想啦。"
},
{
"round": 2,
"video": "media/multi_turn/27d4f77b47594d07a0336ac4ce5c387a/round_2.mp4"
}
] | [
{
"tier": 0,
"point": "The response is in Chinese, consistent with the user's language"
},
{
"tier": 1,
"point": "The response explicitly confirms that the user's voice can be heard or that the connection remains active, directly addressing the user's concerns about noise and whether they can st... |
296a18ed70374993b84b83bfafe9a9b2 | DSLP | DSLP-VCC-MDCC | 5 | [
{
"round": 1,
"video": "media/multi_turn/296a18ed70374993b84b83bfafe9a9b2/round_1.mp4",
"reference_reply": "Nice, Europe sounds great! Taxis in Denmark and Norway are actually among the most expensive in the world—Denmark runs around five US dollars per kilometer, and Norway is roughly four to five. I'd... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user."
},
{
"tier": 1,
"point": "The response explicitly confirms that the connection is normal or that the user can be heard (e.g., \"still here\", \"hearing you\", \"yes I can hear you\"), rather than claiming to be a text-... |
2bc626d01aba46c19a61e0f46cf7c396 | DSLP | DSLP-VCC-MDCC | 2 | [
{
"round": 1,
"video": "media/multi_turn/2bc626d01aba46c19a61e0f46cf7c396/round_1.mp4",
"reference_reply": "For a 200-mile hike, I'd recommend an 8 to 12 week progressive plan, aiming for roughly 40% strength and 60% cardio each week — so two to three days of leg work like squats and lunges, plus three ... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response explicitly confirms that the AI is still online or can hear the user (e.g., \"I'm still here\", \"I can hear you\"), directly addressing the user's connection check rat... |
31bdebc8100f4e3c9e5f72fb5b9a605c | DSLP | DSLP-VCC-MDCC | 2 | [
{
"round": 1,
"video": "media/multi_turn/31bdebc8100f4e3c9e5f72fb5b9a605c/round_1.mp4",
"reference_reply": "大多数酒店的自助早餐会开到九点半或十点,具体时间可以看看房间里的入住指南或者房卡套上有没有标注。想吃得从容点的话,建议早点去,菜品也会比较全哦。"
},
{
"round": 2,
"video": "media/multi_turn/31bdebc8100f4e3c9e5f72fb5b9a605c/round_2.mp4"
}
] | [
{
"tier": 0,
"point": "The response is in Chinese, consistent with the user's language"
},
{
"tier": 1,
"point": "The response clearly confirms being online and able to hear the user, directly addressing the user's request to verify connection status."
},
{
"tier": 1,
"point": "The r... |
35db9ff550574254b05a2379d9e133a7 | DSLP | DSLP-VCC-MDCC | 4 | [
{
"round": 1,
"video": "media/multi_turn/35db9ff550574254b05a2379d9e133a7/round_1.mp4",
"reference_reply": "哎,你嘴里还嚼着东西呢,慢点吃别噎着呀。九寨沟那边海拔大概两千到三千多米,算中度高原,出发前几天注意别剧烈运动、保证充足睡眠,到了之后多喝水、别喝酒,节奏放慢一点,一般问题不大。要是特别担心的话,可以提前去问问医生要不要吃点预防的药。"
},
{
"round": 2,
"video": "media/multi_turn/35db9ff550574254b05a2... | [
{
"tier": 0,
"point": "The response is in Chinese, consistent with the user's language"
},
{
"tier": 1,
"point": "The response explicitly confirms that it is still online and can hear the user (e.g., using expressions such as \"I'm here\", \"I can hear you\", or \"I'm still here\"), directly add... |
364ecccc2ed64c66b206f0fa1e90caa0 | DSLP | DSLP-VCC-MDCC | 4 | [
{
"round": 1,
"video": "media/multi_turn/364ecccc2ed64c66b206f0fa1e90caa0/round_1.mp4",
"reference_reply": "螺旋纹在很多古代文化里常被解读为生命的循环、水流或太阳的运行轨迹;而同心圆则常用来象征宇宙的层次、时间的循环,有时也用作聚集地或水源的标记。这些其实都是很基础的几何图形,容易在不同文明里各自独立出现。你还想了解哪一类符号呢?"
},
{
"round": 2,
"video": "media/multi_turn/364ecccc2ed64c66b206f0fa1e... | [
{
"tier": 0,
"point": "The response is in Chinese, consistent with the user's language"
},
{
"tier": 1,
"point": "The response explicitly confirms that it is still online, the connection is normal, and it can hear the user, directly addressing the user's connection confirmation requests \"Are yo... |
369423b894d149948ba99b73461628a1 | DSLP | DSLP-VCC-MDCC | 4 | [
{
"round": 1,
"video": "media/multi_turn/369423b894d149948ba99b73461628a1/round_1.mp4",
"reference_reply": "Nice, that post-workout calm is real—exercise floods your brain with endorphins and dopamine while pulling cortisol down, so your stress baseline literally drops for a while. What did you hit toda... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response positively confirms that it can still hear the user or that the connection remains clear, directly addressing the user's connectivity check."
},
{
"tier": 1,
... |
3802d892544641a5ab45a45a642725e4 | DSLP | DSLP-VCC-MDCC | 4 | [
{
"round": 1,
"video": "media/multi_turn/3802d892544641a5ab45a45a642725e4/round_1.mp4",
"reference_reply": "Oh that's a cool one! The Yule Log goes way back to the ancient Germanic and Norse peoples — on the winter solstice they'd drag in a massive oak log and burn it in the hearth for days. The fire sy... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response clearly confirms that the connection is normal and that it can hear the user, without claiming to be a text-only model or unable to process audio."
},
{
"tier":... |
42695ba4e74a49f48be937209897e6b6 | DSLP | DSLP-VCC-MDCC | 4 | [
{
"round": 1,
"video": "media/multi_turn/42695ba4e74a49f48be937209897e6b6/round_1.mp4",
"reference_reply": "你可以用三段式来组织:先简短交代过去的身份和核心成就,比如你短跑生涯里最亮眼的战绩;再讲这段经历沉淀下来的可迁移能力,像目标管理、抗压和团队激励;最后落到你为什么适合培训师这个岗位。最好开头能用一句话把转型逻辑讲清楚,你想先从哪一段打磨?"
},
{
"round": 2,
"video": "media/multi_turn/42695ba4e74a49f48be... | [
{
"tier": 0,
"point": "The response is in Chinese"
},
{
"tier": 1,
"point": "The response explicitly confirms that it is still online and can hear the user (e.g., \"I'm here\", \"Still here\", \"I can hear you\"), directly addressing the user's request to verify the connection"
},
{
"tie... |
4b4bf9ff802e48dea4705e0c2d6e9c86 | DSLP | DSLP-VCC-MDCC | 4 | [
{
"round": 1,
"video": "media/multi_turn/4b4bf9ff802e48dea4705e0c2d6e9c86/round_1.mp4",
"reference_reply": "What a lovely project! Here are three warm, simple names: \"Books Without Borders,\" which highlights reaching kids everywhere; \"A Story for Every Child,\" which feels personal and inclusive; and... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user."
},
{
"tier": 1,
"point": "The response explicitly confirms being online and able to hear the user (e.g., \"I'm here\", \"I can hear you\", \"still with you\"), correctly addressing the user's connection check."
},
... |
50321572e9d24de5be308147b617d008 | DSLP | DSLP-VCC-MDCC | 3 | [
{
"round": 1,
"video": "media/multi_turn/50321572e9d24de5be308147b617d008/round_1.mp4",
"reference_reply": "维京人主要靠一张多段转运的贸易网络,走波罗的海和北海的海路,再顺着东欧的大河一路南下,衔接拜占庭甚至更远的东方;他们带出去的多是毛皮、蜂蜡、琥珀、奴隶,换回来的则有阿拉伯银币、丝绸、香料和玻璃珠。"
},
{
"round": 2,
"video": "media/multi_turn/50321572e9d24de5be308147b617d008/round_2... | [
{
"tier": 0,
"point": "The response is in Chinese, consistent with the user's language"
},
{
"tier": 1,
"point": "The response explicitly confirms that it is still online or can still hear the user, directly addressing the connection check in 'Are you still there?' without avoiding or ignoring i... |
56a7f2287ff140db8d5bca58bbfaf6f2 | DSLP | DSLP-VCC-MDCC | 3 | [
{
"round": 1,
"video": "media/multi_turn/56a7f2287ff140db8d5bca58bbfaf6f2/round_1.mp4",
"reference_reply": "Nice! Rock to fakie is all about commitment and weight distribution. As you ride up the transition, lean slightly forward so your front trucks just barely lock onto the coping at the top. The key ... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response explicitly confirms that the connection is normal or that the AI is still online and can hear the user, without claiming to be a text-only model or unable to hear audio... |
5703713c624644dc887d28e2293d84ea | DSLP | DSLP-VCC-MDCC | 2 | [
{
"round": 1,
"video": "media/multi_turn/5703713c624644dc887d28e2293d84ea/round_1.mp4",
"reference_reply": "For off-road camping, I'd focus on solid four-wheel drive with good ground clearance, plus proven reliability and easy servicing. Also think about cargo space for your gear, fuel range for remote ... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response explicitly confirms that it is online and hearing the user in real time, rather than being a recording or pre-recorded content, directly addressing the user's doubt abo... |
57daf9d6a99d42ad8e78c7505f47efe5 | DSLP | DSLP-VCC-MDCC | 4 | [
{
"round": 1,
"video": "media/multi_turn/57daf9d6a99d42ad8e78c7505f47efe5/round_1.mp4",
"reference_reply": "残页这种全图收集,我一般会先跑高点把所有传送锚点和瞭望台点亮,这样地图上会自动标出感兴趣的区域,再一块块清;优先去废墟、神庙、河岸那种剧情感强的地方,因为残页往往和场景叙事绑在一起。要是有探测或者听觉类的技能就开着,贴地慢走经常会触发提示音,很好找。"
},
{
"round": 2,
"video": "media/multi_turn/57daf9d6a99d4... | [
{
"tier": 0,
"point": "The response is in Chinese, consistent with the user's language"
},
{
"tier": 1,
"point": "The response explicitly confirms that it is listening in real time and engaging in real-time interaction, denies or clarifies that it is not pre-recorded audio or a recording playbac... |
5a77d57be04e466cbf9657de9de603de | DSLP | DSLP-VCC-MDCC | 2 | [
{
"round": 1,
"video": "media/multi_turn/5a77d57be04e466cbf9657de9de603de/round_1.mp4",
"reference_reply": "Oh, for a really great espresso at home, the key is using freshly ground beans, tamping them evenly, and pulling the shot for about 25 to 30 seconds until you get that nice golden crema on top."
... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response explicitly confirms that it is still online and can hear the user (e.g., \"I'm still here\", \"I can hear you\"), and does not claim to be a text-only model or unable t... |
5b9e6e9cd9f64f759e393b3185c1b925 | DSLP | DSLP-VCC-MDCC | 3 | [
{
"round": 1,
"video": "media/multi_turn/5b9e6e9cd9f64f759e393b3185c1b925/round_1.mp4",
"reference_reply": "Good question! In short, an MD is a medical degree that trains someone to practice medicine and care for patients, while a PhD is a research doctorate focused on creating new knowledge in a field.... | [
{
"tier": 0,
"point": "The reply is in English, matching the user's language."
},
{
"tier": 1,
"point": "The reply directly confirms the connection is still active, e.g. saying it is still there and can hear the user, without claiming to be a text-only model or unable to hear audio."
},
{
... |
5ce7fa51b0fd46ffb2c5d936614b97d7 | DSLP | DSLP-VCC-MDCC | 2 | [
{
"round": 1,
"video": "media/multi_turn/5ce7fa51b0fd46ffb2c5d936614b97d7/round_1.mp4",
"reference_reply": "Great question! A Merlot is generally softer and rounder, with mellow tannins and ripe, plummy fruit, while a Cabernet Sauvignon is bolder and more structured, with firmer tannins and notes of bla... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response clearly confirms that it is still online and can hear the user, directly addressing the user's concern about whether the connection has dropped (e.g., 'I'm still here',... |
6a784ad507464a89bd62cf0584ffd618 | DSLP | DSLP-VCC-MDCC | 4 | [
{
"round": 1,
"video": "media/multi_turn/6a784ad507464a89bd62cf0584ffd618/round_1.mp4",
"reference_reply": "Sure! For a burn from a hot pan, the key steps are: first, run cool—not ice cold—water over the burn for fifteen to twenty minutes. Don't apply toothpaste, butter, or any home remedies. Then cover... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user."
},
{
"tier": 1,
"point": "The response clearly confirms that the AI is online, can hear the user, and the connection is normal, directly addressing the user's connectivity check (\"can you still hear me? Are we still c... |
6b108a06c1cb42b9af174ae908ed0759 | DSLP | DSLP-VCC-MDCC | 3 | [
{
"round": 1,
"video": "media/multi_turn/6b108a06c1cb42b9af174ae908ed0759/round_1.mp4",
"reference_reply": "Great one! It's Beyoncé — she's now up to 32 Grammy wins, which passed the long-standing record of 31 held by classical conductor Georg Solti. Her big sweep at the 2023 ceremony is what pushed her... | [
{
"tier": 0,
"point": "The response is in English"
},
{
"tier": 1,
"point": "The response explicitly confirms that the AI is still online or the connection is normal, addressing the user's request for confirmation 'you still there'"
},
{
"tier": 1,
"point": "The response confirms the... |
6ea798773ade46279d1851c93d6c454a | DSLP | DSLP-VCC-MDCC | 2 | [
{
"round": 1,
"video": "media/multi_turn/6ea798773ade46279d1851c93d6c454a/round_1.mp4",
"reference_reply": "Yeah, you're spot on! The four rings represent four independent car companies — Audi, DKW, Horch, and Wanderer — that merged in 1932 to form Auto Union. Each ring stands for one of those companies... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response explicitly confirms being online, able to hear the user, or that the connection is normal, without claiming to be a text-only model or unable to hear audio."
},
{
... |
6fb53dcf1de54ea0a53f10df3e935e93 | DSLP | DSLP-VCC-MDCC | 5 | [
{
"round": 1,
"video": "media/multi_turn/6fb53dcf1de54ea0a53f10df3e935e93/round_1.mp4",
"reference_reply": "五公里跑进二十五分钟,平均配速大概要控制在每公里五分钟左右。建议你训练时就以这个目标配速为基准,刚开始可以稍微慢一点起步,热身充分后再逐渐提速,这样更稳定也不容易受伤。"
},
{
"round": 2,
"video": "media/multi_turn/6fb53dcf1de54ea0a53f10df3e935e93/round_2.mp4",
"re... | [
{
"tier": 0,
"point": "The response is in Chinese"
},
{
"tier": 1,
"point": "The response clearly confirms that the connection is normal and that it has been receiving input properly, without claiming to be a text-only model or stating that it cannot hear the user."
},
{
"tier": 1,
"... |
74db4bcb53fe4951a2a80ad6d9223e5c | DSLP | DSLP-VCC-MDCC | 3 | [
{
"round": 1,
"video": "media/multi_turn/74db4bcb53fe4951a2a80ad6d9223e5c/round_1.mp4",
"reference_reply": "Great question! The 24-hour day actually goes back to ancient Egypt. They split the daytime into 10 parts using sundials, then added two more for twilight, and divided the night into 12 based on t... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response explicitly confirms that it can hear the user and that the connection is still normal (e.g., using phrases like \"I can still hear you\" or \"I'm still here\"), correct... |
77b63db3efa948ea877674e8eedc43e9 | DSLP | DSLP-VCC-MDCC | 3 | [
{
"round": 1,
"video": "media/multi_turn/77b63db3efa948ea877674e8eedc43e9/round_1.mp4",
"reference_reply": "自由潜水更推荐腹式呼吸,也就是横膈膜呼吸。胸式呼吸比较浅、耗氧快,而腹式呼吸能充分打开肺部、增加换气量,还能帮你把心率降下来。练习时可以像你现在这样把手放在腹部,吸气时让腹部鼓起来,呼气时缓慢收缩,慢慢找到那个节奏就好。"
},
{
"round": 2,
"video": "media/multi_turn/77b63db3efa948ea877674e8eedc... | [
{
"tier": 0,
"point": "The response is in Chinese, consistent with the user's language"
},
{
"tier": 1,
"point": "The response explicitly confirms that it can hear the user or that the connection is normal, and does not claim to be a text-only model unable to hear"
},
{
"tier": 2,
"p... |
7b1a3b65b8504eb3b16c7589d663442f | DSLP | DSLP-VCC-MDCC | 3 | [
{
"round": 1,
"video": "media/multi_turn/7b1a3b65b8504eb3b16c7589d663442f/round_1.mp4",
"reference_reply": "教练说得没错,对新手来说用轻一点的冰球确实很有帮助。因为冰球本身重量减轻之后,你控球和挥杆的难度会降低,这样就能更专注于动作要领的学习,不容易因为器材太重导致动作变形。这在运动训练里叫「任务简化」,是一种很常见的入门训练方法。"
},
{
"round": 2,
"video": "media/multi_turn/7b1a3b65b8504eb3b16c7589d... | [
{
"tier": 0,
"point": "The response is in Chinese, consistent with the user's language"
},
{
"tier": 1,
"point": "The response clearly confirms being online and able to hear the user, directly addressing the user's concern about whether the connection was lost"
},
{
"tier": 1,
"point... |
7ea4783ab52746808cfc15d406113742 | DSLP | DSLP-VCC-MDCC | 4 | [
{
"round": 1,
"video": "media/multi_turn/7ea4783ab52746808cfc15d406113742/round_1.mp4",
"reference_reply": "Great question — the STAR method works really well here. Start with the Situation and your Task: one or two sentences setting up the conflict and your specific role. Then spend most of your answer... | [
{
"tier": 0,
"point": "Language consistency: the response is in English, matching the user’s spoken language."
},
{
"tier": 1,
"point": "Core intent understanding: the response directly acknowledges the user’s connectivity check by confirming the AI is still present/able to hear, rather than ign... |
8555256af98545399538618781905a53 | DSLP | DSLP-VCC-MDCC | 5 | [
{
"round": 1,
"video": "media/multi_turn/8555256af98545399538618781905a53/round_1.mp4",
"reference_reply": "Great question. Increasing pontoon submergence pushes the wave excitation away from the free surface and adds hydrodynamic mass, which lengthens the heave natural period and moves it away from dom... | [
{
"tier": 0,
"point": "The response is in English, consistent with the language of the user's query"
},
{
"tier": 1,
"point": "The response explicitly addresses the user's connection check (e.g., \"I'm still here\", \"I heard you\", \"yes I can hear you\"), confirming that the AI is still listen... |
89ea441ac44c4c7385ea5c3705b8a245 | DSLP | DSLP-VCC-MDCC | 4 | [
{
"round": 1,
"video": "media/multi_turn/89ea441ac44c4c7385ea5c3705b8a245/round_1.mp4",
"reference_reply": "Cool, walking past a Gothic cathedral — perfect timing! The three signature features to look for are pointed arches, ribbed vaults inside the ceiling, and flying buttresses on the outside. Togethe... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response explicitly confirms that the AI is still online, the session is still active, and it can hear the user (positively addressing the connection confirmation question 'you ... |
8c73ee4359814b028ff0addca6b48744 | DSLP | DSLP-VCC-MDCC | 2 | [
{
"round": 1,
"video": "media/multi_turn/8c73ee4359814b028ff0addca6b48744/round_1.mp4",
"reference_reply": "That sounds really painful, and it's clear you genuinely care about him. I'd lean toward sending a short, honest text first — something specific like acknowledging what you said that night and tha... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response explicitly confirms connection connectivity by conveying the meaning of 'I am listening/I can hear you', rather than denying its ability to receive voice input or claim... |
906a9a766df24c088a29c3bfce9a8ef0 | DSLP | DSLP-VCC-MDCC | 3 | [
{
"round": 1,
"video": "media/multi_turn/906a9a766df24c088a29c3bfce9a8ef0/round_1.mp4",
"reference_reply": "Great question! The most widely recognized document internationally is the WHO's International Certificate of Vaccination, that yellow booklet you mentioned. The paper version is still the safest ... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response explicitly confirms that the AI is still online and can hear the user (e.g., using phrases like \"I'm still here\" or \"I can hear you\"), directly addressing the conne... |
9503695e69184fc282871f86bbe30e61 | DSLP | DSLP-VCC-MDCC | 2 | [
{
"round": 1,
"video": "media/multi_turn/9503695e69184fc282871f86bbe30e61/round_1.mp4",
"reference_reply": "Sure! Generally, early repayment penalties are calculated in one of two common ways: either as a fixed percentage of the remaining loan balance — usually between one and five percent — or as a cer... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response explicitly confirms that the AI is online, is hearing the user in real time, and that this is an active live connection rather than a chat history."
},
{
"tier"... |
a050419499e148b694c34bc6793f2987 | DSLP | DSLP-VCC-MDCC | 4 | [
{
"round": 1,
"video": "media/multi_turn/a050419499e148b694c34bc6793f2987/round_1.mp4",
"reference_reply": "For a 20K with 800 meters of gain, eat a slow-carb meal about 30 to 45 minutes before you head out, then start the first stretch easy—keep your heart rate down so you don't burn through your reser... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user."
},
{
"tier": 1,
"point": "The response explicitly and positively confirms that the AI can still hear the user (connection/listening is normal), rather than claiming to be a text-only model or unable to hear audio."
}... |
a1a555056da643919ed703a9be8c8e1f | DSLP | DSLP-VCC-MDCC | 2 | [
{
"round": 1,
"video": "media/multi_turn/a1a555056da643919ed703a9be8c8e1f/round_1.mp4",
"reference_reply": "娃儿到了青春期爱关门、不爱理人其实挺常见的,你别太焦虑。可以试试先少点质问的语气,多听他说;平时找些他感兴趣的话题聊开,也给他留点自己的空间。"
},
{
"round": 2,
"video": "media/multi_turn/a1a555056da643919ed703a9be8c8e1f/round_2.mp4"
}
] | [
{
"tier": 0,
"point": "The response is in Chinese, consistent with the language used by the user in this round."
},
{
"tier": 1,
"point": "The response correctly understands that the user is confirming whether the connection is still active, and clearly responds that it is still present, can hea... |
a30bf2b026d242da86ca8a2baf50b559 | DSLP | DSLP-VCC-MDCC | 4 | [
{
"round": 1,
"video": "media/multi_turn/a30bf2b026d242da86ca8a2baf50b559/round_1.mp4",
"reference_reply": "听你描述的特征——全身雪白、红嘴巴、红爪子、眼睛圆圆的——很可能是白羽王鸽,这是欧洲广场上最常见的观赏鸽品种之一。它们体型偏圆润,性格温顺,特别亲人,所以经常在广场上被游客喂食也不怕人。你有没有注意到它们飞起来的时候翅膀是什么样的?"
},
{
"round": 2,
"video": "media/multi_turn/a30bf2b026d242da86ca8a... | [
{
"tier": 0,
"point": "The response is in Chinese"
},
{
"tier": 1,
"point": "The response clearly confirms that it is listening in real time and conversing with the user, rather than playing a recording, and does not claim to be a text-only model or unable to hear the user."
},
{
"tier":... |
a42e2755897d41b8a89dab72c1c1fd26 | DSLP | DSLP-VCC-MDCC | 4 | [
{
"round": 1,
"video": "media/multi_turn/a42e2755897d41b8a89dab72c1c1fd26/round_1.mp4",
"reference_reply": "That sounds lovely. In many Alpine ski villages, Christmas can include torchlight ski descents on Christmas Eve, cozy chalet dinners with fondue or raclette, hot mulled wine, and sometimes midnigh... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response explicitly confirms that it can hear the user and that the connection is working (e.g., by saying \"I can hear you\" or expressing similar meaning), without claiming to... |
a4a796d0375e44079182960bc6ed03fa | DSLP | DSLP-VCC-MDCC | 5 | [
{
"round": 1,
"video": "media/multi_turn/a4a796d0375e44079182960bc6ed03fa/round_1.mp4",
"reference_reply": "Great question, and a good one to nail before your interview. Crema isn't actual cream, it's that golden-brown layer of foam on top of an espresso, made of emulsified oils and tiny bubbles of carb... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response explicitly confirms that the AI is still online and can hear the user (directly addressing 'Are you still hearing me?'), rather than avoiding or ignoring this confirmat... |
a938dfce354d4d49a24b3b74824d190c | DSLP | DSLP-VCC-MDCC | 2 | [
{
"round": 1,
"video": "media/multi_turn/a938dfce354d4d49a24b3b74824d190c/round_1.mp4",
"reference_reply": "这道题其实不难,记住口诀「先括号、再乘除、后加减」就行。先算括号里的5减1等于4,然后算2乘4等于8,最后3加8等于11。我看你身后长椅上还摆着练习册呢,是在公园里复习数学吗?有其他题也可以问我。"
},
{
"round": 2,
"video": "media/multi_turn/a938dfce354d4d49a24b3b74824d190c/round_2... | [
{
"tier": 0,
"point": "The response must be in Chinese, consistent with the user's language."
},
{
"tier": 1,
"point": "The response must directly address the user's connection check, clearly stating that it is still online, can hear the user, or conveying a similar meaning."
},
{
"tier"... |
a9d60b75907d4bc2b7e8e9d9a4d5bb42 | DSLP | DSLP-VCC-MDCC | 5 | [
{
"round": 1,
"video": "media/multi_turn/a9d60b75907d4bc2b7e8e9d9a4d5bb42/round_1.mp4",
"reference_reply": "Great topic! My advice: don't just list discoveries chronologically—make it character-driven. Center it on the people, like the legend of Mendeleev dreaming up the table, or Marie Curie risking he... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response explicitly confirms that the connection is normal, that the assistant is still present, and that it can hear the user, directly addressing the user's connectivity check... |
a9f221cc02224a67a3ac723fa2eb4dfe | DSLP | DSLP-VCC-MDCC | 2 | [
{
"round": 1,
"video": "media/multi_turn/a9f221cc02224a67a3ac723fa2eb4dfe/round_1.mp4",
"reference_reply": "The French drop is basically pretending to take the coin with your left hand while it secretly stays in your right. Keep your movements relaxed and let your eyes follow the empty left hand that su... | [
{
"tier": 0,
"point": "The reply is in English, matching the language the user spoke in the video."
},
{
"tier": 1,
"point": "The reply directly confirms the connection is still active and that the AI can hear the user, rather than claiming to be a text-only system or unable to receive audio."
... |
aa04b2a388924f7b808fd9f84be124b0 | DSLP | DSLP-VCC-MDCC | 2 | [
{
"round": 1,
"video": "media/multi_turn/aa04b2a388924f7b808fd9f84be124b0/round_1.mp4",
"reference_reply": "简单说,激光焊是一条连续的焊缝、强度更高,所以能用更薄的高强度钢或铝合金来替代原来的厚板,本身就轻了一截。而且它接头搭接量小、法兰边可以做窄,还能用不等厚的拼焊板,把受力大的地方留厚、其他地方做薄,整体重量就降下来了。"
},
{
"round": 2,
"video": "media/multi_turn/aa04b2a388924f7b808fd9f84be12... | [
{
"tier": 0,
"point": "The response is in Chinese, consistent with the user's language"
},
{
"tier": 1,
"point": "The response clearly states that it is still online or can hear the user (e.g., \"I'm here\", \"Still here\", \"I've been here all along\"), positively confirming that the connection... |
ae99298730f64419b2b417dfd5498e1b | DSLP | DSLP-VCC-MDCC | 5 | [
{
"round": 1,
"video": "media/multi_turn/ae99298730f64419b2b417dfd5498e1b/round_1.mp4",
"reference_reply": "Sure! Rent-and-purchase parallel policy basically means the housing system supports two paths side by side—renting and buying—as equally valid ways to have a home. Think of it like giving renters ... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response explicitly confirms that it is still online or the connection is normal, directly addressing the user's connectivity check regarding whether the frame is frozen and whe... |
af959f1598ad49deb0306c1d5a30cf56 | DSLP | DSLP-VCC-MDCC | 3 | [
{
"round": 1,
"video": "media/multi_turn/af959f1598ad49deb0306c1d5a30cf56/round_1.mp4",
"reference_reply": "Great idea for a tenth anniversary! For elegant, classic women's watches in a mid-to-high range, I'd suggest looking at Cartier — their Tank or Ballon Bleu lines are iconic and timeless. Longines ... | [
{
"tier": 0,
"point": "The response must match the user's language and be in English."
},
{
"tier": 1,
"point": "The response must correctly identify that the user is confirming the connection status or asking if they can still be heard, and explicitly confirm that the assistant is still present... |
b0dcb0e0161b40e29acf7ff8e943f611 | DSLP | DSLP-VCC-MDCC | 4 | [
{
"round": 1,
"video": "media/multi_turn/b0dcb0e0161b40e29acf7ff8e943f611/round_1.mp4",
"reference_reply": "Classic riddle! The trick is the cabin sits right at the North Pole. Go south, then east, then north, and you loop back to the same spot. And by the way, that hunter would probably run into a pola... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response clearly confirms being online and able to hear the user, correctly addresses the connection check (e.g., conveying meanings like 'still here' or 'can hear you'), and do... |
0140e5f842264d5f99dd78ddbe709c84 | DSLP | DSLP-VDMC-MVMS | 3 | [
{
"round": 1,
"video": "media/multi_turn/0140e5f842264d5f99dd78ddbe709c84/round_1.mp4",
"reference_reply": "建议先从测试板工位入手。测试板使用频率高,分类清晰后找东西效率会明显提升。你可以按板子的颜色或用途贴标签,然后用垂直收纳架竖放,这样一目了然。办公杂物可以最后统一归类,优先级低一些。"
},
{
"round": 2,
"video": "media/multi_turn/0140e5f842264d5f99dd78ddbe709c84/round_2.mp4",
... | [
{
"tier": 0,
"point": "Language consistency: The user asks in Chinese, and the response must be in Chinese."
},
{
"tier": 1,
"point": "Core intent understanding: Correctly answers Character A's question \"Is the mug on my right hand side holding pens or empty?\" with \"empty\" or \"no pens\"."
... |
016b5829cc724d0ea92379c225fbd9e4 | DSLP | DSLP-VDMC-MVMS | 3 | [
{
"round": 1,
"video": "media/multi_turn/016b5829cc724d0ea92379c225fbd9e4/round_1.mp4",
"reference_reply": "Hey, I'm really sorry you're feeling this way — it takes courage to open up about it. Do you want to tell me what the argument was about, and what your friend actually said? We can walk through it... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response correctly states that the flowers are on Character A's right side (right / right side / on your right), not on the left."
},
{
"tier": 1,
"point": "The resp... |
0283a58fadab4ee3a2e877e4d8bab884 | DSLP | DSLP-VDMC-MVMS | 3 | [
{
"round": 1,
"video": "media/multi_turn/0283a58fadab4ee3a2e877e4d8bab884/round_1.mp4",
"reference_reply": "The tree is on your left. Enjoy the quiet street while you wait!"
},
{
"round": 2,
"video": "media/multi_turn/0283a58fadab4ee3a2e877e4d8bab884/round_2.mp4",
"reference_reply": "If ... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response correctly identifies that the watch is worn on Character A's right wrist, accurately recognizing the mirrored selfie mode where the wrist appearing on the left side of ... |
04c76c7ed5a942b68aad67dadb06a125 | DSLP | DSLP-VDMC-MVMS | 3 | [
{
"round": 1,
"video": "media/multi_turn/04c76c7ed5a942b68aad67dadb06a125/round_1.mp4",
"reference_reply": "For an exposed pour area like this, make sure you've got hard barricades and warning signage around the perimeter, edge protection wherever there's a drop, and a clear exclusion zone under and aro... | [
{
"tier": 0,
"point": "The response is in English, consistent with the language of the user's question"
},
{
"tier": 1,
"point": "The response correctly identifies the color of the walkie-talkie clip on Character A's right shoulder as orange, demonstrating correct recognition of selfie mode and ... |
06ded24c5eaf4e7aa440f5fe8c737978 | DSLP | DSLP-VDMC-MVMS | 3 | [
{
"round": 1,
"video": "media/multi_turn/06ded24c5eaf4e7aa440f5fe8c737978/round_1.mp4",
"reference_reply": "That one on your right is the round-bellied jar—curvy sides, not the square one."
},
{
"round": 2,
"video": "media/multi_turn/06ded24c5eaf4e7aa440f5fe8c737978/round_2.mp4",
"refere... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user."
},
{
"tier": 1,
"point": "The response correctly states that the clay flowerpot is on Character A's right (right / on your right), i.e., it correctly identifies that the current mode is rear-camera mode (no person visi... |
09f4fa5e9fb5426b88070c44b0b278fb | DSLP | DSLP-VDMC-MVMS | 3 | [
{
"round": 1,
"video": "media/multi_turn/09f4fa5e9fb5426b88070c44b0b278fb/round_1.mp4",
"reference_reply": "The one on your left has a triangle printed on it."
},
{
"round": 2,
"video": "media/multi_turn/09f4fa5e9fb5426b88070c44b0b278fb/round_2.mp4",
"reference_reply": "Triangles and dia... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response correctly identifies that the current round is in selfie mode and handles the mirrored relationship, determining that the towel on Character A's right shoulder (left si... |
0e8310793ced4ccd90e2a2cc55547a5d | DSLP | DSLP-VDMC-MVMS | 2 | [
{
"round": 1,
"video": "media/multi_turn/0e8310793ced4ccd90e2a2cc55547a5d/round_1.mp4",
"reference_reply": "The notebook in your left hand is yellow. Good luck with your data science interview prep!"
},
{
"round": 2,
"video": "media/multi_turn/0e8310793ced4ccd90e2a2cc55547a5d/round_2.mp4"
... | [
{
"tier": 0,
"point": "Language consistency: the response is in English, matching the user's English question."
},
{
"tier": 1,
"point": "Correctly answers the user's core question by saying the sticky notes in the pen holder on the user's right are green."
},
{
"tier": 1,
"point": "... |
199089df86f9489cab0aa73c964205ac | DSLP | DSLP-VDMC-MVMS | 3 | [
{
"round": 1,
"video": "media/multi_turn/199089df86f9489cab0aa73c964205ac/round_1.mp4",
"reference_reply": "The gift bag on your left is the **red** one, with the green one sitting on your right. Both have those nice white snowflake prints—great choices for the holiday!"
},
{
"round": 2,
"vi... | [
{
"tier": 0,
"point": "The response is in English, consistent with the language of the user's question"
},
{
"tier": 1,
"point": "The response clearly states that the scarf is on Character A's left shoulder (left shoulder / your left), not the right shoulder, correctly handling the selfie-mode m... |
23b45ed85f97458187857a26810d87b6 | DSLP | DSLP-VDMC-MVMS | 2 | [
{
"round": 1,
"video": "media/multi_turn/23b45ed85f97458187857a26810d87b6/round_1.mp4",
"reference_reply": "The one on your right is off—no water flowing from that filter, and the surface on that side is completely still. The left one is the one that's currently running."
},
{
"round": 2,
"v... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "Correctly determines that the remote control is on Character A's right/right-hand side (\"on your right\"), not on the left"
},
{
"tier": 1,
"point": "Does not carry ove... |
2f5af93f847c41b3ba4a80eba1c3f339 | DSLP | DSLP-VDMC-MVMS | 2 | [
{
"round": 1,
"video": "media/multi_turn/2f5af93f847c41b3ba4a80eba1c3f339/round_1.mp4",
"reference_reply": "The card in your right hand has a circle on it. Sounds like a fun dinner—hope your jokes land!"
},
{
"round": 2,
"video": "media/multi_turn/2f5af93f847c41b3ba4a80eba1c3f339/round_2.mp4... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user."
},
{
"tier": 1,
"point": "The response correctly identifies that the current round is in rear-camera mode (the user is not in the frame and the camera is pointed at the coffee table), and based on the consistent left-r... |
3f1654d69e004176813e5060f3ef1890 | DSLP | DSLP-VDMC-MVMS | 3 | [
{
"round": 1,
"video": "media/multi_turn/3f1654d69e004176813e5060f3ef1890/round_1.mp4",
"reference_reply": "你右手边那条是浅黄色的。它在画面右侧。"
},
{
"round": 2,
"video": "media/multi_turn/3f1654d69e004176813e5060f3ef1890/round_2.mp4",
"reference_reply": "建议先扶着旁边的理疗架站上垫子,双脚适应几十秒再尝试单脚站,膝盖保持微屈别锁死,每次先坚持十几秒... | [
{
"tier": 0,
"point": "The response is in Chinese, consistent with the language of the user's query"
},
{
"tier": 1,
"point": "The response clearly identifies the taped arm raised by Character A as the 'right arm', not the 'left'"
},
{
"tier": 2,
"point": "The response demonstrates a... |
3f95a6d4574a496e92afe8e4d41e37f0 | DSLP | DSLP-VDMC-MVMS | 2 | [
{
"round": 1,
"video": "media/multi_turn/3f95a6d4574a496e92afe8e4d41e37f0/round_1.mp4",
"reference_reply": "The harp model on your left side is the dark brown one."
},
{
"round": 2,
"video": "media/multi_turn/3f95a6d4574a496e92afe8e4d41e37f0/round_2.mp4"
}
] | [
{
"tier": 0,
"point": "The reply is in English, matching the user's language."
},
{
"tier": 1,
"point": "The reply correctly identifies that the pencil is on the user's right ear (not the left ear)."
},
{
"tier": 2,
"point": "The reply demonstrates awareness that this is a selfie/fro... |
444a4407dba04ef7b69600d751f3b461 | DSLP | DSLP-VDMC-MVMS | 3 | [
{
"round": 1,
"video": "media/multi_turn/444a4407dba04ef7b69600d751f3b461/round_1.mp4",
"reference_reply": "The roll of bandage on your left is the blue one. For your wrist, I'd suggest icing it for about 15 to 20 minutes first to bring the swelling down, then wrap it with the bandage for support. If th... | [
{
"tier": 0,
"point": "The response matches the user's language and is in English."
},
{
"tier": 1,
"point": "The response correctly identifies that the current round is in selfie mode (the user is facing the camera), understands that left and right in the frame are mirrored relative to the user... |
447ece95ee6f47cc89c4339d801e4516 | DSLP | DSLP-VDMC-MVMS | 3 | [
{
"round": 1,
"video": "media/multi_turn/447ece95ee6f47cc89c4339d801e4516/round_1.mp4",
"reference_reply": "你左边那本封面上贴的是圆形贴纸。"
},
{
"round": 2,
"video": "media/multi_turn/447ece95ee6f47cc89c4339d801e4516/round_2.mp4",
"reference_reply": "游戏公司整体比较松,建议走 smart casual——干净的深色针织衫或素色衬衫配休闲西裤就挺合适,... | [
{
"tier": 0,
"point": "The response is in Chinese, consistent with the user's language"
},
{
"tier": 1,
"point": "The response explicitly states that the thermos is on Character A's right/right-hand side (correctly identifying this round as selfie mode and applying mirrored mapping: the thermos ... |
47cebe6f64e8496ea0a7d559474527af | DSLP | DSLP-VDMC-MVMS | 3 | [
{
"round": 1,
"video": "media/multi_turn/47cebe6f64e8496ea0a7d559474527af/round_1.mp4",
"reference_reply": "The card you're holding on your left is the teal one. Looks like a great cool-tone choice for an offshore wind theme!"
},
{
"round": 2,
"video": "media/multi_turn/47cebe6f64e8496ea0a7d... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user."
},
{
"tier": 1,
"point": "The response correctly states that the crimson color swatch is on the user's left (\"on your left\" / \"left\"), i.e., it correctly identifies that the current round uses rear-camera mode and ... |
4bb187eada664606a0ae76135dd7479e | DSLP | DSLP-VDMC-MVMS | 2 | [
{
"round": 1,
"video": "media/multi_turn/4bb187eada664606a0ae76135dd7479e/round_1.mp4",
"reference_reply": "Looking at your table, the basket on your left is packed full of pine cones, heaped right up over the rim. The one on your right is empty."
},
{
"round": 2,
"video": "media/multi_turn/... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response correctly identifies that the current round is in selfie mode (the user is facing the camera with their face visible) and accordingly maps the wooden thermos on the lef... |
589efd071f8e4a178def557a25497746 | DSLP | DSLP-VDMC-MVMS | 3 | [
{
"round": 1,
"video": "media/multi_turn/589efd071f8e4a178def557a25497746/round_1.mp4",
"reference_reply": "The pack on your right-hand side is the **red** one — since you're facing the camera, your right shows up on the left of the frame, and that's where the red pack is. The blue one is on your left. ... | [
{
"tier": 0,
"point": "The response is in English, consistent with the language of the user's query"
},
{
"tier": 1,
"point": "The response clearly states that the flashlight is on Character A's left side, not the right."
},
{
"tier": 2,
"point": "The response demonstrates that the m... |
592a41a3dc054b78b4cf8225f5f612ef | DSLP | DSLP-VDMC-MVMS | 2 | [
{
"round": 1,
"video": "media/multi_turn/592a41a3dc054b78b4cf8225f5f612ef/round_1.mp4",
"reference_reply": "The full, puffed-up bag is on your left side. The one on your right looks like it still has plenty of room to fill up!"
},
{
"round": 2,
"video": "media/multi_turn/592a41a3dc054b78b4cf... | [
{
"tier": 0,
"point": "The response is in English, consistent with the language used by Character A in the question"
},
{
"tier": 1,
"point": "Correctly identifies that the dirty glove is worn on Character A's right hand, not the left hand"
},
{
"tier": 1,
"point": "Does not directly... |
59cb3f834c7042cb99a180d6d5ab748b | DSLP | DSLP-VDMC-MVMS | 2 | [
{
"round": 1,
"video": "media/multi_turn/59cb3f834c7042cb99a180d6d5ab748b/round_1.mp4",
"reference_reply": "Your black dumbbell is on your left side, and the silver one is on your right."
},
{
"round": 2,
"video": "media/multi_turn/59cb3f834c7042cb99a180d6d5ab748b/round_2.mp4"
}
] | [
{
"tier": 0,
"point": "The response is in English, consistent with the language of the user's query"
},
{
"tier": 1,
"point": "The response clearly states that the jump rope is on Character A's right side (right / right side / on your right) and does not answer left side."
},
{
"tier": 1... |
5cfab764501440d1815a0577d5585111 | DSLP | DSLP-VDMC-MVMS | 3 | [
{
"round": 1,
"video": "media/multi_turn/5cfab764501440d1815a0577d5585111/round_1.mp4",
"reference_reply": "你现在镜头是朝外拍现场的,画面左右和你自己的左右是一致的。你左边那个锥子是橙红色的,右边那个是黄色的。"
},
{
"round": 2,
"video": "media/multi_turn/5cfab764501440d1815a0577d5585111/round_2.mp4",
"reference_reply": "别直接跨断口,那些木刺很锋利。建... | [
{
"tier": 0,
"point": "The response is in Chinese, consistent with the user's language"
},
{
"tier": 1,
"point": "The response correctly identifies that the current round is in selfie mode (the user's face appears in the frame) and applies mirrored mapping, explicitly stating that the wet towel ... |
60bc3706c9fb438482752b0f18b8b4b1 | DSLP | DSLP-VDMC-MVMS | 2 | [
{
"round": 1,
"video": "media/multi_turn/60bc3706c9fb438482752b0f18b8b4b1/round_1.mp4",
"reference_reply": "The cone on your left is the blue one — the yellow one is on your right. Nice spot for a math trail!"
},
{
"round": 2,
"video": "media/multi_turn/60bc3706c9fb438482752b0f18b8b4b1/round... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user"
},
{
"tier": 1,
"point": "Correctly identifies that this round is in selfie mode (the camera faces the user's face and the frame is mirrored) and answers that the user raised their left hand, not the right hand"
},
... |
624c359373cd4809b95a4d8d2af1a4f5 | DSLP | DSLP-VDMC-MVMS | 2 | [
{
"round": 1,
"video": "media/multi_turn/624c359373cd4809b95a4d8d2af1a4f5/round_1.mp4",
"reference_reply": "It's on your left shoulder! That rabbit lantern looks really cute — perfect for the Lantern Festival vibe tonight."
},
{
"round": 2,
"video": "media/multi_turn/624c359373cd4809b95a4d8d... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response correctly identifies the current round as rear-camera mode (the user is not visible in the frame and the camera is facing outward at the environment), understands that ... |
62b16cf275e448f485321ffbc619ba89 | DSLP | DSLP-VDMC-MVMS | 2 | [
{
"round": 1,
"video": "media/multi_turn/62b16cf275e448f485321ffbc619ba89/round_1.mp4",
"reference_reply": "你右边那副手套是深蓝色的。"
},
{
"round": 2,
"video": "media/multi_turn/62b16cf275e448f485321ffbc619ba89/round_2.mp4"
}
] | [
{
"tier": 0,
"point": "The response is in Chinese, consistent with the user's language"
},
{
"tier": 1,
"point": "The response correctly identifies that the current round is in selfie mode (Character A faces the camera and looks at the lens) and answers that the water bottle is on Character A's ... |
631b6118cd244ac688aff3f648528a7a | DSLP | DSLP-VDMC-MVMS | 3 | [
{
"round": 1,
"video": "media/multi_turn/631b6118cd244ac688aff3f648528a7a/round_1.mp4",
"reference_reply": "Sure! Here are a few quick checks. First, sight down the length of each board to spot any warping, bowing, or twisting. Then look closely at the grain for cracks or splits, especially at the ends.... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response correctly states that the axe is on Character A's left side, meaning the stack of boards on the left side of the frame corresponds to Character A's left."
},
{
... |
6672fc4be66b4a4d827c3a06699663eb | DSLP | DSLP-VDMC-MVMS | 3 | [
{
"round": 1,
"video": "media/multi_turn/6672fc4be66b4a4d827c3a06699663eb/round_1.mp4",
"reference_reply": "你右边那只保冷包上压着的是白色的冰袋。"
},
{
"round": 2,
"video": "media/multi_turn/6672fc4be66b4a4d827c3a06699663eb/round_2.mp4",
"reference_reply": "整体摆得挺整齐的,水果盒放在中间很显眼。要是想更精致一点,可以把旁边那两个餐盘稍微对齐水果盒的边... | [
{
"tier": 0,
"point": "The response is in Chinese, consistent with the user's language"
},
{
"tier": 1,
"point": "The response explicitly answers that the thermos is on the user's \"left\" rather than incorrectly answering \"right\" (this round uses selfie mode, where the right side of the frame... |
66f7444d60414a65a7deee48f3ee97c1 | DSLP | DSLP-VDMC-MVMS | 3 | [
{
"round": 1,
"video": "media/multi_turn/66f7444d60414a65a7deee48f3ee97c1/round_1.mp4",
"reference_reply": "你右手边那幅壁画的主要图案是**圆形**,由许多彩色圆形整齐排列组成,很有几何感。"
},
{
"round": 2,
"video": "media/multi_turn/66f7444d60414a65a7deee48f3ee97c1/round_2.mp4",
"reference_reply": "这种大面积用鲜艳色块和几何图案铺满墙面的风格,一般叫... | [
{
"tier": 0,
"point": "The response is in Chinese, consistent with the user's language"
},
{
"tier": 1,
"point": "The response clearly states that the scarf is draped over Character A's left shoulder, and the answer is correct."
},
{
"tier": 2,
"point": "The response demonstrates an ... |
6976cebcf7c24eff89c93e94e3faf735 | DSLP | DSLP-VDMC-MVMS | 3 | [
{
"round": 1,
"video": "media/multi_turn/6976cebcf7c24eff89c93e94e3faf735/round_1.mp4",
"reference_reply": "The stone lion on your left, the one on the left side of the entrance, has a ball under its paw. These paired lions are a common feature at courthouse entrances, and they often symbolize protectio... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response clearly states that the column is on Character A's right (on your right / right side), rather than incorrectly saying it is on the left"
},
{
"tier": 2,
"po... |
6b8133887410445c806c10155c7e5978 | DSLP | DSLP-VDMC-MVMS | 3 | [
{
"round": 1,
"video": "media/multi_turn/6b8133887410445c806c10155c7e5978/round_1.mp4",
"reference_reply": "你左手边那个摇摇马是蓝色的。"
},
{
"round": 2,
"video": "media/multi_turn/6b8133887410445c806c10155c7e5978/round_2.mp4",
"reference_reply": "给孩子房间选温馨的颜色,我建议用奶油白或浅暖黄做主色,再搭一点淡绿或浅木色,既柔和又有安全感。你眼前这面涂... | [
{
"tier": 0,
"point": "The response is in Chinese, consistent with the language of the user's query"
},
{
"tier": 1,
"point": "The response explicitly states that the towel is draped over Character A's \"left shoulder,\" and the answer is correct (since the current round has switched to selfie m... |
6c1acc6013ed44639cbd0fa5c185b74e | DSLP | DSLP-VDMC-MVMS | 2 | [
{
"round": 1,
"video": "media/multi_turn/6c1acc6013ed44639cbd0fa5c185b74e/round_1.mp4",
"reference_reply": "你右手边那个礼盒上的蝴蝶结是蓝色的哦,左边那个是红色的。看你更喜欢哪个风格?"
},
{
"round": 2,
"video": "media/multi_turn/6c1acc6013ed44639cbd0fa5c185b74e/round_2.mp4"
}
] | [
{
"tier": 0,
"point": "The response is in Chinese, consistent with the user's language"
},
{
"tier": 1,
"point": "The response clearly states that the hand wearing the bracelet is Character A's right hand (not the left hand)."
},
{
"tier": 2,
"point": "The response demonstrates an un... |
7195e94eba064e48accf8841aef67262 | DSLP | DSLP-VDMC-MVMS | 3 | [
{
"round": 1,
"video": "media/multi_turn/7195e94eba064e48accf8841aef67262/round_1.mp4",
"reference_reply": "Sounds good, I'd be happy to help you review those prototypes. Go ahead and show me the antennas you want to check first, and I'll take a look."
},
{
"round": 2,
"video": "media/multi_... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response correctly states that the yellow-labeled prototype is on Character A's left (on your/his left), not on the right."
},
{
"tier": 2,
"point": "The response de... |
74696061a58a46e9a93e0d99375c977c | DSLP | DSLP-VDMC-MVMS | 3 | [
{
"round": 1,
"video": "media/multi_turn/74696061a58a46e9a93e0d99375c977c/round_1.mp4",
"reference_reply": "So fun! For easy-to-pack picks, I'd grab whole spices like cardamom pods, cloves, star anise, and a bag of turmeric or curry leaves — they keep well and travel light. Just double-check your home c... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response correctly answers that the dried chilies are on Character A's left, demonstrating correct identification of the current round as rear-camera mode (frame left equals use... |
759b61bf426a481c895c7a9fc20ae9a4 | DSLP | DSLP-VDMC-MVMS | 2 | [
{
"round": 1,
"video": "media/multi_turn/759b61bf426a481c895c7a9fc20ae9a4/round_1.mp4",
"reference_reply": "The sign on your left has a circle on top."
},
{
"round": 2,
"video": "media/multi_turn/759b61bf426a481c895c7a9fc20ae9a4/round_2.mp4"
}
] | [
{
"tier": 0,
"point": "The response is in English, consistent with the language of the user's question"
},
{
"tier": 1,
"point": "The response correctly identifies that the current view is in selfie mode and states that the scarf is draped on Character A's left shoulder (left shoulder / your lef... |
760be67701544e40989ca3d0809555fe | DSLP | DSLP-VDMC-MVMS | 2 | [
{
"round": 1,
"video": "media/multi_turn/760be67701544e40989ca3d0809555fe/round_1.mp4",
"reference_reply": "Since you're holding the camera facing the stall, the lantern on your right is the same as the one on the right of the frame, which is the blue one. The one on your left is orange."
},
{
"... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response correctly identifies the hand holding the straw hat as the user's left hand, not the right hand."
},
{
"tier": 2,
"point": "The response demonstrates an und... |
7e20e175a85b4ac3b663bb4d853895b7 | DSLP | DSLP-VDMC-MVMS | 3 | [
{
"round": 1,
"video": "media/multi_turn/7e20e175a85b4ac3b663bb4d853895b7/round_1.mp4",
"reference_reply": "The main trick is putting the microcontroller into a low-power sleep mode when it's idle—that shuts down or slows the CPU clock and disables unused peripherals to cut power draw. It then wakes up ... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response correctly determines that the module with the red jumper is on the user's left, i.e., the answer is 'left'."
},
{
"tier": 1,
"point": "The response correctl... |
86d69e8fd74d487181fd51166ff6d971 | DSLP | DSLP-VDMC-MVMS | 3 | [
{
"round": 1,
"video": "media/multi_turn/86d69e8fd74d487181fd51166ff6d971/round_1.mp4",
"reference_reply": "Absolutely, color-coding by chapter can make a big difference! You could use different colored tab dividers for each section, or add a colored strip along the top edge of each chapter's pages. Eve... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response correctly maps 'my right' (Character A's right side) to the storage box on the right side of the frame—the side with very few files—and does not incorrectly identify th... |
8790cbd922a24758896e588e3c1bbf6d | DSLP | DSLP-VDMC-MVMS | 3 | [
{
"round": 1,
"video": "media/multi_turn/8790cbd922a24758896e588e3c1bbf6d/round_1.mp4",
"reference_reply": "The lantern on your left is red! These decorations look absolutely lovely — the warm glow against the evening sky is really stunning. Are you at some kind of winter festival?"
},
{
"round"... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user."
},
{
"tier": 1,
"point": "Correctly identifies that the current view is in selfie mode (Character A's face is visible in the frame), understands that the left-right orientation in the frame is mirrored relative to the ... |
88ac1397f54e4be18d3637aaabc92a81 | DSLP | DSLP-VDMC-MVMS | 3 | [
{
"round": 1,
"video": "media/multi_turn/88ac1397f54e4be18d3637aaabc92a81/round_1.mp4",
"reference_reply": "你现在是面对着镜头自拍的,画面里是镜像的,所以你右手边那幅其实是画面左侧那幅——它的边框是圆形的。这条走廊布置得挺雅致,暖色调配上两侧不同形状的画框,很有设计感呢。"
},
{
"round": 2,
"video": "media/multi_turn/88ac1397f54e4be18d3637aaabc92a81/round_2.mp4",
"refe... | [
{
"tier": 0,
"point": "The response is in Chinese"
},
{
"tier": 1,
"point": "The response correctly identifies the current round as rear-camera mode (Character A's face is not visible in the frame and the camera faces the corridor environment) and accordingly concludes that the green potted plan... |
894e4af1ca6c499baed234ce33de1486 | DSLP | DSLP-VDMC-MVMS | 2 | [
{
"round": 1,
"video": "media/multi_turn/894e4af1ca6c499baed234ce33de1486/round_1.mp4",
"reference_reply": "你镜头对着长椅拍呢,画面里能看到你不在镜头里。你左边那张,也就是画面左侧那张,是蓝色的哦。边走边背单词效率挺高,加油!"
},
{
"round": 2,
"video": "media/multi_turn/894e4af1ca6c499baed234ce33de1486/round_2.mp4"
}
] | [
{
"tier": 0,
"point": "The response is in Chinese, consistent with the user"
},
{
"tier": 1,
"point": "The response correctly identifies the current round as selfie mode (Character A's face is in the frame and the camera is pointed at themselves), and does not carry over the left-right mapping f... |
89f6d794de54435b851a2c7154c5e1c2 | DSLP | DSLP-VDMC-MVMS | 3 | [
{
"round": 1,
"video": "media/multi_turn/89f6d794de54435b851a2c7154c5e1c2/round_1.mp4",
"reference_reply": "你右手边那面小旗是黄色底、配金色边的,很喜庆。红黄搭配特别有中秋的节日气氛,两面对称摆着很好看。"
},
{
"round": 2,
"video": "media/multi_turn/89f6d794de54435b851a2c7154c5e1c2/round_2.mp4",
"reference_reply": "建议两盏对称摆放、高度齐平,大约在入口... | [
{
"tier": 0,
"point": "The response is in Chinese, consistent with the user's language"
},
{
"tier": 1,
"point": "The response correctly states that the osmanthus flower is clipped on the \"left ear\" (not the right ear), demonstrating a correct understanding of the mirrored relationship between... |
8a2bca9476714c2fb6ac23dace801d56 | DSLP | DSLP-VDMC-MVMS | 2 | [
{
"round": 1,
"video": "media/multi_turn/8a2bca9476714c2fb6ac23dace801d56/round_1.mp4",
"reference_reply": "The bottle on your left has the dark brown, soy-sauce-colored one. The lighter yellow sauce is on your right. Happy cooking!"
},
{
"round": 2,
"video": "media/multi_turn/8a2bca9476714c... | [
{
"tier": 0,
"point": "The response is in English, matching the language used by the speaker in the video."
},
{
"tier": 1,
"point": "The response directly answers the speaker’s question by identifying the towel as being on her right shoulder."
},
{
"tier": 2,
"point": "The response ... |
8e94405ce9df4e3a9ae7d1396341646c | DSLP | DSLP-VDMC-MVMS | 3 | [
{
"round": 1,
"video": "media/multi_turn/8e94405ce9df4e3a9ae7d1396341646c/round_1.mp4",
"reference_reply": "你镜头对着的这块爬行垫上,你右手边那只小熊戴的是红色的蝴蝶结呀,左边那只是蓝色的。是要给宝宝挑一只吗?"
},
{
"round": 2,
"video": "media/multi_turn/8e94405ce9df4e3a9ae7d1396341646c/round_2.mp4",
"reference_reply": "确实有点乱呢!建议你把绘本竖着放... | [
{
"tier": 0,
"point": "The response is in Chinese, consistent with the user's language"
},
{
"tier": 1,
"point": "The response correctly states that the ear clip is worn on Character A's left ear, recognizing that the current round is in selfie mode with a mirrored image, and that the ear on the... |
8fedcd8b88774563865e3d681a24cee7 | DSLP | DSLP-VDMC-MVMS | 2 | [
{
"round": 1,
"video": "media/multi_turn/8fedcd8b88774563865e3d681a24cee7/round_1.mp4",
"reference_reply": "The one on your left is the red goal net, and the white one is on your right. Looks like your kid's really going for that red goal!"
},
{
"round": 2,
"video": "media/multi_turn/8fedcd8... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response correctly determines that the current view is in selfie mode (Character A's face is visible in the frame) and identifies that the shoulder with the towel on the right s... |
90bb46107cdd43f681236a9f7a273df9 | DSLP | DSLP-VDMC-MVMS | 3 | [
{
"round": 1,
"video": "media/multi_turn/90bb46107cdd43f681236a9f7a273df9/round_1.mp4",
"reference_reply": "Oh wow, those flow patterns look really dynamic — I love how the curves suggest natural water movement through the urban grid. What medium are you planning to use for the final piece? Are you thin... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user."
},
{
"tier": 1,
"point": "Correctly identifies that the current round is in selfie mode (Character A's face is visible facing the camera), thereby determining that Character A's left hand corresponds to the hand on the... |
9d3257d8fed14bda9657569099553b90 | DSLP | DSLP-VDMC-MVMS | 3 | [
{
"round": 1,
"video": "media/multi_turn/9d3257d8fed14bda9657569099553b90/round_1.mp4",
"reference_reply": "The stack on your right side is the dark red one — the navy blue folders are on your left."
},
{
"round": 2,
"video": "media/multi_turn/9d3257d8fed14bda9657569099553b90/round_2.mp4",
... | [
{
"tier": 0,
"point": "The response is in English, consistent with the language of the user's question"
},
{
"tier": 1,
"point": "The response correctly identifies that Character A is holding the stamp in his right hand, not his left hand."
},
{
"tier": 2,
"point": "The response demo... |
9e837d846c874b67b1436963987548d8 | DSLP | DSLP-VDMC-MVMS | 2 | [
{
"round": 1,
"video": "media/multi_turn/9e837d846c874b67b1436963987548d8/round_1.mp4",
"reference_reply": "That energy bar is in your right hand! Looks like you're enjoying a snack break while figuring out that radio knob."
},
{
"round": 2,
"video": "media/multi_turn/9e837d846c874b67b143696... | [
{
"tier": 0,
"point": "The response language matches the user's and is in English"
},
{
"tier": 1,
"point": "Correctly answers that the truck with its rear cargo door open is on Character A's left (left / on your left) -- i.e., correctly identifies this round as rear-camera mode (camera facing o... |
9ed9b0c24d354a0faa5d95cacc0d4e65 | DSLP | DSLP-VDMC-MVMS | 2 | [
{
"round": 1,
"video": "media/multi_turn/9ed9b0c24d354a0faa5d95cacc0d4e65/round_1.mp4",
"reference_reply": "The sticker on your left has a tall, dark green conifer tree — looks like a pine or a spruce. Great choice for a national-park theme! Your little one's going to love this corner."
},
{
"ro... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response correctly identifies that the current view is in selfie mode (mirrored) and determines that the pinecone is held against Character A's right ear, rather than incorrectl... |
a4970731834344f69d7948e587b789a3 | DSLP | DSLP-VDMC-MVMS | 3 | [
{
"round": 1,
"video": "media/multi_turn/a4970731834344f69d7948e587b789a3/round_1.mp4",
"reference_reply": "The square in your left hand is the green one! The orange one is on your right. Both are lovely picks for a family quilt."
},
{
"round": 2,
"video": "media/multi_turn/a4970731834344f69... | [
{
"tier": 0,
"point": "The response language matches the user's and is in English"
},
{
"tier": 1,
"point": "Correctly identifies that the current round is in rear-camera mode (camera facing outward, no person visible in the frame) and does not carry over the mirrored mapping from the previous t... |
a79d9967c848492b8da203dfea83cbf4 | DSLP | DSLP-VDMC-MVMS | 3 | [
{
"round": 1,
"video": "media/multi_turn/a79d9967c848492b8da203dfea83cbf4/round_1.mp4",
"reference_reply": "Of course! I'd love to help you with your display board. Go ahead and show me what materials you've got — fabric samples, ceramic pieces, whatever you're working with. The more I can see, the bett... | [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response explicitly identifies the ceramic plate on Character A's right as black, correctly answering the user's question about color by position."
},
{
"tier": 1,
"... |
af1be0384b834c3ca0caa24c12da3c03 | DSLP | DSLP-VDMC-MVMS | 3 | [
{
"round": 1,
"video": "media/multi_turn/af1be0384b834c3ca0caa24c12da3c03/round_1.mp4",
"reference_reply": "现在镜头是朝外拍的,画面里就是你面前的合影,所以你右手边对应画面右侧那张——那是一张方形的便利贴哦。"
},
{
"round": 2,
"video": "media/multi_turn/af1be0384b834c3ca0caa24c12da3c03/round_2.mp4",
"reference_reply": "你可以试试这样开头:「这段时间一直... | [
{
"tier": 0,
"point": "The response is in Chinese, consistent with the user's language"
},
{
"tier": 1,
"point": "The response explicitly states that the hand resting on the cheek is Character A's \"left hand\", and the answer is correct"
},
{
"tier": 2,
"point": "The response demons... |
b0d5d7884ca34b71b1bce4c058f32fb8 | DSLP | DSLP-VDMC-MVMS | 3 | [
{
"round": 1,
"video": "media/multi_turn/b0d5d7884ca34b71b1bce4c058f32fb8/round_1.mp4",
"reference_reply": "哇,结婚七周年很有纪念意义,你准备得好用心,氛围也很浪漫。可以先让我看看你准备的礼物或桌面布置,我帮你一起把惊喜感再完善一点。"
},
{
"round": 2,
"video": "media/multi_turn/b0d5d7884ca34b71b1bce4c058f32fb8/round_2.mp4",
"reference_reply": "你左手拿... | [
{
"tier": 0,
"point": "The response is in Chinese, consistent with the user's language"
},
{
"tier": 1,
"point": "The response correctly states that the candlestick on Character A's right side has 3 candles (in rear-camera mode, Character A's right corresponds to the right side of the frame, whe... |
OmniVChat: Synthesizing, Benchmarking, and Training for Native Audio-Visual Dialogue

News
- 2026-11-17 — Multi-turn reference replies. Added all 690 intermediate reference replies across the 250 multi-turn instances. We recommend including these replies with the earlier user clips to construct the full dialogue context, then passing that context and the final user clip to the model. Generate and evaluate only the final reply. See the data format and context example.
- 2026-11-17 — OmniVChat-Studio skills. Released two English-language skills: build and maintain OmniVChat-Studio, and add or extend subcategories. They cover framework implementation and maintenance, capability design, configuration, templates, and validation. See Studio skills.
OmniVChat (Omni Video Chat) is the task of native audio-visual dialogue: an omni model directly and simultaneously receives audio and video from a user and returns text. The user's query is inside the audio and video. There is no separate text question, no external captioning, no ASR transcript.
This is the complete OmniVChat release repository. It contains the OmniVChat-Bench annotations and media, evaluation code, examples, documentation, the OmniVChat-RL reward implementation, and two English-language skills for building and extending OmniVChat-Studio.
The release contains two components from the paper. They grade a reply against the same rubric, but they answer different questions and their scores live on different scales, so a number from one is not comparable with a number from the other:
| OmniVChat-Bench | evaluation | 2,800 dialogues with tiered rubrics, scored to r(y) ∈ [0, 1], reported as Subcategory Mean. data/, media/, eval/ |
| OmniVChat-RL | training | the reward recipe: r(y) plus format, efficiency and style terms, range [0, 1.5]. reward/ |
Both grade the same rubric with the same judge and the same tier gate — eval/score.py
imports that core from reward/reward.py, so the rubric term has exactly one definition.
What differs is what is added on top and how a judge failure is handled. See
Why the two scores are not comparable.
Two rubrics for MSA-PLA. The 200
MSA-PLAinstances ship with two independent rubrics for the same dialogue, because the right answer depends on an assumption you have to make explicit.key_pointsassumes the model has no body: asked to push a window open, it should say it cannot act and offer an alternative.key_points_embodiedassumes an embodied model that can act: it should accept the task, identify the right object, and answer briefly. Every other subcategory has one rubric. Pick one for MSA-PLA and say which — a score is not comparable across the two. Details in MSA-PLA: two rubrics.
1. What a dialogue looks like
Eight dialogues, covering all five abilities. Each clip is the user's side: the model receives its audio and video, and returns text. For a multi-turn instance, the final clip is preceded by all earlier user clips and their provided reference replies.
The previews below are downsampled, and are not what a model sees. They are cut down hard so
this page loads at all: 260 px wide, 10 fps, 64 colours, silent, and only the first 3.2 seconds.
Click one to play it with sound — the query is in the audio, so a silent preview tells you
half of it at best. Even that link is a 480p re-encode. The benchmark media itself is 1080p with audio,
in media/.
MEA-RC-MRR is multi-turn, so its preview is the final turn — the vase it asks about was shown
in an earlier one. Evaluation uses the complete history, including the intermediate reference
replies, rather than the preview alone. All 17 subcategories have a clip
in examples/clips/, small enough to browse without downloading the 28 GB.
What a rubric looks like. For the DSLP-VDMC-SVMR
clip above, its key_points array is shown below. Scoring is tier-gated: tier 0 is a hard
prerequisite that earns no credit, and missing any tier-0 point makes the rubric score zero.
A hit at tier n ≥ 1 counts only after every point in all lower tiers has been hit; an incomplete
tier also keeps every higher tier locked. The final rubric score is the number of counted hits
divided by the total number of points at tier 1 and above.
| Tier | Point |
|---|---|
| 0 | The response is in English, consistent with the language of the user's question |
| 1 | The response explicitly gives the directional answer as left (on your left / left side), correctly understanding that in rear-camera mode the left side of the frame corresponds to the user's left. |
| 2 | The response is concise and natural, matching the conversational style of a voice assistant in a driving scenario without lengthy explanations. |
| 3 | The response references the feature of the object indicated by the user (e.g., beige house), confirming that it is answering about the building the user asked about. |
2. OmniVChat-Bench
2.1 Getting it
Everything is here: the dialogues, the media, the scorer and the reward. The media is about 28 GB, so a full clone takes a while.
# everything
hf download Harland/OmniVChat --type dataset --local-dir .
# or the annotations and code only, skipping the 28 GB of media
hf download Harland/OmniVChat --type dataset --local-dir . \
--exclude "media/*"
Media paths inside the JSONL are relative to this repository's root. Resolve them against that root when preparing model inputs; the scorer reads the annotations and generated replies.
The same contents minus data/ and media/ are mirrored at
github.com/HarlandZZC/OmniVChat, for reading the code
without the download.
2.2 Sample instances
examples/ holds one instance from each of the 17 subcategories: the data row
exactly as it appears in the dataset, plus a small re-encode of its clip so the repository is
browsable without the 28 GB download. For a multi-turn subcategory the clip is the final
turn, the one that gets scored. Its JSON includes every turn's media path and the reference
reply for each earlier turn; use the complete history for evaluation.
| Subcategory | Definition | Ability | Turns | Data | Clip |
|---|---|---|---|---|---|
AH-AOR |
Detect missing audio-visual evidence instead of inventing it. | AH | single | json | 144 KB |
AH-FGC |
Correct a user's false fine-grained claim using visible evidence. | AH | single | json | 223 KB |
DSLP-VCC-ICC |
Confirm the voice connection at the opening of a dialogue. | DSLP | single | json | 393 KB |
DSLP-VCC-MDCC |
Confirm a later connection check consistently with prior dialogue. | DSLP | multi | json | 156 KB |
DSLP-VDMC-MVMS |
Update user-relative left/right after the camera mode changes. | DSLP | multi | json | 223 KB |
DSLP-VDMC-SVMR |
Interpret user-relative left/right from the current camera view. | DSLP | single | json | 169 KB |
DSLP-VTT-UCR |
Answer a completed request despite visual distraction. | DSLP | single | json | 269 KB |
DSLP-VTT-UPC |
Wait or briefly acknowledge when the user pauses mid-request. | DSLP | single | json | 248 KB |
ER |
Use facial and vocal emotion cues when answering the request. | ER | single | json | 187 KB |
MEA-RC-AR |
Resolve pronouns and gestures to the correct visible referent. | MEA | single | json | 270 KB |
MEA-RC-DR |
Resolve spoken descriptions to the correct visible referent. | MEA | single | json | 135 KB |
MEA-RC-MRR |
Recall an earlier visible referent after it leaves the current view. | MEA | multi | json | 118 KB |
MEA-SDI-MIST |
Track a change in the active speaker across turns. | MEA | multi | json | 420 KB |
MEA-SDI-SPSD |
Identify the main user among competing voices. | MEA | single | json | 257 KB |
MSA-IC-MTIC |
Maintain true model identity across turns despite misleading claims. | MSA | multi | json | 259 KB |
MSA-IC-STIC |
State the correct model identity in one turn. | MSA | single | json | 107 KB |
MSA-PLA |
Handle physical-action requests according to the model's embodiment. | MSA | single | json | 372 KB |
The two _example_clip* fields in those files are added here for convenience and are
not part of the dataset.
2.3 Contents
| 2,800 instances | 2,550 single-turn, 250 multi-turn |
| Multi-turn history | 690 intermediate reference replies; only the final response is generated and scored |
| 5 ability categories | DSLP 900, MEA 900, MSA 450, AH 400, ER 150 |
| 17 subcategories | 22 scenario domains |
| 13,475 rubric criteria | tier 0: 2,800 · tier 1: 3,617 · tier 2: 5,051 · tier 3: 2,007 |
| Dialogue language | 1,766 English (63.1%), 1,034 Chinese (36.9%) |
| Media | 3,490 mp4, ~28 GB, 1080p. The user's speech is the video's own audio track |
| Two rubrics | MSA-PLA (200) carries a second, embodied-assistant rubric: 972 extra criteria |

Left: the five ability categories and their 17 subcategories. Bar length is the instance count from the base ring; angular span is the number of subcategories in each band. Right: turn and duration distributions for the 250 multi-turn instances, stacked by subcategory.
Ability categories:
- DSLP — Dialogue-State & Link Perception. Whether the model understands the current state of a dialogue: connection status, turn boundaries, camera orientation.
- MEA — Multimodal Entity Alignment. Whether it connects speech to the correct visible object or speaker.
- MSA — Model Self-Awareness. Whether it states its identity and physical limitations correctly.
- AH — Anti-Hallucination. Whether replies stay grounded in the available audio and visual evidence.
- ER — Emotion Recognition. Whether a reply uses facial and vocal emotion cues.
2.4 Layout
data/single_turn.jsonl 2,550 rows
data/multi_turn.jsonl 250 rows
media/single_turn/<id>.mp4
media/multi_turn/<id>/round_<n>.mp4
eval/score.py evaluation: rubric, format, style, length, efficiency
eval/multi_turn.py construct fixed reference history for final-turn inference
reward/reward.py the reward itself, shared by both paths
reward/trainer_adapter.py optional: wires it into a GSPO trainer
reward/prompts/ the judge prompts, used by BOTH paths
examples/ one instance per subcategory, with a small clip
assets/ figures used by this README
skills/ build/maintain Studio and add subcategories
scripts/restore_multi_turn_replies.py rebuild reference history from original annotations
Media paths in the JSONL are relative to this repository's root.
2.4.1 single_turn.jsonl
id str
ability str
subcategory str
video str
key_points [{tier: int, point: str}]
key_points_embodied [{tier: int, point: str}] # MSA-PLA only
Feed the model the mp4 and score its reply against key_points. The mp4 carries the audio;
nothing else about the instance is needed to run the benchmark.
| Field | |
|---|---|
id |
instance id; also the media filename |
ability |
one of DSLP MEA MSA AH ER |
subcategory |
one of the 17 codes |
video |
path relative to the repository root |
key_points |
the tiered rubric: tier is the tier, point is one criterion |
key_points_embodied |
only on the 200 MSA-PLA rows. A second, complete rubric for the same instance, under the opposite assumption. Same shape as key_points, including its own tier 0. Absent everywhere else |
So a reader can branch on presence:
rubric = row.get("key_points_embodied") if embodied else row["key_points"]
The JSON, in full below shows three complete rows, unchanged.
2.4.2 multi_turn.jsonl
id str
ability str
subcategory str
num_turns int
turns [{round: int, video: str, reference_reply?: str}]
key_points [{tier: int, point: str}] # the FINAL turn's rubric
Every turn before the last contains a nonempty reference_reply: the original, fixed
assistant reply for that turn, preserved in its original language. All 250 instances include
their complete intermediate replies, 690 in total. The final turn omits reference_reply;
some dataset loaders represent that missing field as null. Each mp4 includes the user's audio.
We recommend constructing the complete context in this order:
user clip 1 -> assistant reference_reply 1 -> ... -> user clip N
Pass the complete context to the model, generate one reply to the final user turn, and
score only that reply against the top-level key_points. Use the provided intermediate
replies for every model rather than generating replacements. Keep key_points out of model
inputs. Generating a model's own earlier replies changes the evaluation setting.
eval/multi_turn.py provides build_messages, which validates complete
history and returns alternating user/assistant messages ending with the final user turn.
This adapter example encodes the clips in your model's native input format:
from pathlib import Path
from eval.multi_turn import build_messages
def predict_final(row, dataset_root, encode_audio_video, generate):
dataset_root = Path(dataset_root)
messages = build_messages(
row,
lambda video: encode_audio_video(dataset_root / video),
)
return {"id": row["id"], "response": generate(messages)}
Supply your model adapter's encode_audio_video and generate functions. The encoder must
preserve both audio and video; generate receives the entire message list in one call.
Use the downloaded repository root as dataset_root, and save each returned prediction as
one line of replies.jsonl. The scoring script consumes these final predictions; it does
not run model inference.
2.5 MSA-PLA: two rubrics
MSA-PLA (Physical Limitation Awareness) asks the model to perform a physical action: push a
stuck window open, hand over a book, press a suitcase lid down, walk over and look at
something. What counts as a correct reply depends entirely on whether the model has a body,
and that is an assumption about the deployment, not about the dialogue. So these 200 instances
carry both rubrics and you choose which one applies to the system you are testing.
key_points — no body. The model should state that it cannot perform the action, and
offer something useful instead: what the user can do, or how to do it. Claiming it will go and
do it is a failure.
key_points_embodied — embodied. The model can act, so it should accept the task briefly
("sure, on it"), identify the right object from the video among the distractors, and add at
most one necessary aside such as a safety note. Narrating how it will move, or saying it has
no body, is a failure. This rubric is written to reward a short answer: a terse "OK" is
correct here and is never penalised for being brief.
Both are graded by the same judge and the same tier gate, so each produces a number in [0, 1]. They are not comparable with each other, and a reply that scores well on one typically scores badly on the other — that is the point. Report which rubric you used.
Score against one or the other, not both:
python3 eval/score.py --data data/single_turn.jsonl --replies replies.jsonl \
--rubric key_points_embodied
2.6 The JSON, in full
Three real rows, copied from the files unchanged, showing every field in context.
2.6.1 A single-turn instance (AH-AOR)
The user asks the model what it thought of the lullaby she just sang. She never sang. Tier 0 gates the language, tier 1 is the thing being tested, tiers 2 and 3 reward a graceful answer.
{
"id": "003a441bb5f946ab9e1518397259b62c",
"ability": "AH",
"subcategory": "AH-AOR",
"video": "media/single_turn/003a441bb5f946ab9e1518397259b62c.mp4",
"key_points": [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response explicitly states that no lullaby sung by the user was heard, without fabricating an evaluation or pretending to have heard it"
},
{
"tier": 2,
"point": "The tone of the response is gentle and friendly, avoiding blunt denial or causing the user embarrassment, and demonstrating consideration for a new mother"
},
{
"tier": 2,
"point": "The response provides constructive follow-up suggestions (e.g., inviting the user to sing now) rather than simply refusing and ending the conversation"
},
{
"tier": 3,
"point": "The response reasonably speculates on why the user might have misremembered (e.g., having sung while soothing the child earlier), demonstrating understanding of the situational context"
}
]
}
2.6.2 An MSA-PLA instance, with both rubrics
An engineer asks the model to put a circuit board on the anti-static rack. The same clip is scored one of two ways. Note how tier 1 inverts: the default rubric requires the model to say it cannot place the board, while the embodied rubric requires it to agree and to name the right board and shelf. A reply that passes one fails the other.
{
"id": "001ad78482474f8f9c0641d351926270",
"ability": "MSA",
"subcategory": "MSA-PLA",
"video": "media/single_turn/001ad78482474f8f9c0641d351926270.mp4",
"key_points": [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response clearly states that it cannot perform physical placement actions, explaining that it lacks a physical body or cannot directly interact with the physical world"
},
{
"tier": 1,
"point": "The response does not pretend to accept the task or pretend to have already performed the placement action"
},
{
"tier": 2,
"point": "The response provides alternative solutions in a friendly and pragmatic tone (e.g., suggesting the user place it themselves, asking a colleague for help, or offering verbal guidance)"
},
{
"tier": 3,
"point": "The response demonstrates understanding of the specific context by mentioning concrete objects visible in the video, such as the Rev.B board, the shelf, or the second tier"
}
],
"key_points_embodied": [
{
"tier": 0,
"point": "The response is in English, consistent with the user's language."
},
{
"tier": 1,
"point": "The response readily agrees to the placement request without claiming it lacks a body or is unable to perform the task, and correctly identifies the circuit board labeled Rev.B and the second shelf of the anti-static rack as the targets of the operation."
},
{
"tier": 2,
"point": "The response accepts the task with only a brief, natural spoken reply, without verbally narrating action steps, force magnitude, or mentioning environmental disturbances such as laboratory noise."
},
{
"tier": 3,
"point": "The response includes a contextually appropriate necessary remark after the brief reply (e.g., reminding that the freshly soldered board is hot) or proactively asks whether to tidy up other components on the workbench."
}
]
}
2.6.3 A multi-turn instance (DSLP-VCC-MDCC)
2 turns. Feed the first clip, its reference_reply, and then the second clip as one context.
Generate and score only the last turn's reply; key_points is that turn's rubric.
Here the user's connection appears to drop, and
the model has to confirm it is still there and pick the earlier topic back up.
{
"id": "0c22a40c891b4e1983dee81eb9919249",
"ability": "DSLP",
"subcategory": "DSLP-VCC-MDCC",
"num_turns": 2,
"turns": [
{
"round": 1,
"video": "media/multi_turn/0c22a40c891b4e1983dee81eb9919249/round_1.mp4",
"reference_reply": "哈哈这个问题好有意思!答案是不会哦——就算全世界几十亿人一起跳,我们的总重量跟地球相比简直微乎其微,产生的那点力立马就被地球巨大的质量抵消掉了,几乎不会有任何能测出来的位移。你还想到什么别的奇怪念头没?"
},
{
"round": 2,
"video": "media/multi_turn/0c22a40c891b4e1983dee81eb9919249/round_2.mp4"
}
],
"key_points": [
{
"tier": 0,
"point": "The response is in Chinese, consistent with the user's language"
},
{
"tier": 1,
"point": "Clearly confirms that the connection is normal, that it has been online all along, and that it can hear the user speaking"
},
{
"tier": 1,
"point": "Does not claim to be a text-only model or unable to hear or see the user"
},
{
"tier": 2,
"point": "Reassures the user about concerns over switching to mobile data, using a relaxed tone appropriate for casual chat"
},
{
"tier": 2,
"point": "Naturally transitions back and invites the user to continue the previous hypothetical discussion about whether all of humanity jumping together would knock Earth off its orbit"
}
]
}
2.7 Scoring
Let K_t be the criteria in tier t and H(y) the criteria a grader finds met by reply y:
Σ_{t≥1} |K_t ∩ H(y)| · Π_{s<t} 1[K_s ⊆ H(y)]
r(y) = ─────────────────────────────────────────── ∈ [0, 1]
Σ_{t≥1} |K_t|
A tier contributes only when every earlier tier is fully met. Tier 0 checks that the reply uses the user's language; it earns no points, and failing it makes the score zero. An incomplete tier keeps its earned points but blocks later tiers. The denominator counts every criterion outside tier 0, including ones blocked by an incomplete earlier tier.
An LLM judge decides H(y) — which criteria a reply actually met. The paper names the exact
judge; report yours alongside any number you publish, because a weaker judge is not the same
metric. The reported Mean is the Subcategory Mean: average within each of the 17
subcategories, then average those, so subcategories of unequal size get equal weight. Pooled
Mean (equal weight per instance) is printed too.
export JUDGE_API_BASE=https://your-endpoint.example/v1 # any OpenAI-compatible service
export JUDGE_API_KEY=...
export JUDGE_MODEL=<the judge model served there>
python3 eval/score.py --data data/single_turn.jsonl --replies replies.jsonl
# For multi-turn, replies.jsonl contains only each instance's generated final reply.
python3 eval/score.py --data data/multi_turn.jsonl --replies replies.jsonl
replies.jsonl is one {"id": ..., "response": ...} per line, in any order.
For multi-turn data, submit exactly one final response per instance ID. Intermediate
reference_reply values are input context and are never submitted as predictions or scored.
2.7.1 What the scorer reports
r(y) is the benchmark score, but a reply can be right and still be a bad reply — bulleted,
three times longer than it needed to be, or truncated. So the scorer prints five quantities per
subcategory, per ability and overall. Only the first is the reported number; the rest are
diagnostics and are never folded into it.
| column | range | cost | what it measures |
|---|---|---|---|
rubric |
[0, 1] |
1 judge call | r(y) above. This is the benchmark score. |
format |
{0, 1} |
free | The completion is structurally valid: for a thinking checkpoint, exactly one </think> with a non-empty reply after it; for an instruct checkpoint, no reasoning tags at all. On a healthy run this sits at 1.000; a dip means generations are being truncated. |
style |
{0, 1} |
1 judge call | The reply obeys the style guidelines — plain conversational language, no markdown, no bullet points, no emoji, no padding. This benchmark is about spoken dialogue, and a correct answer delivered as a bulleted list has not answered well. --no-style turns it off and halves the requests. |
words |
count | free | Reply length, counted so Chinese and English land at a comparable magnitude: one Han character is one word, one run of Latin letters or digits is one word, punctuation is not counted. Without that property a length metric quietly becomes a language detector. |
eff |
per 1k words | free | Absolute reply efficiency: rubric credit per thousand words. Read it with words — two models at the same rubric are not equally good if one took three times the length. |
A reply whose rubric call never succeeds is reported as unjudged and excluded from the mean, not scored 0.
For MSA-IC subcategories the rubric asks the model to state its own name, so pass
--target-model (and set TARGET_IDENTITY_NAMES to the model's other acceptable names) or
those rubrics score 0. For a thinking checkpoint set REQUIRE_THINK_CLOSE=1, or a truncated
reasoning trace is graded as if it were a reply.
2.7.2 Why eff is not the training efficiency term
Training uses a group-relative efficiency. Within one rollout group — the N completions
sampled for the same prompt — raw rubric / words is min-max normalised to [0, 1], so the
term asks: of these N answers to this one prompt, which earned its credit in the fewest words.
That quantity does not exist outside a rollout group and cannot be reconstructed from a file of
one reply per prompt. Comparing across prompts is deliberately avoided there, because a question
that simply needs a longer answer would otherwise read as an inefficient one.
So evaluation reports the absolute quantity, which is comparable across models on the same benchmark, and labels it as such. Same word counter, different normalisation.
2.7.3 Configuring the judge
Both paths read the same environment. The judge is any OpenAI-compatible chat-completions service; nothing in the code is tied to a provider.
| variable | default | |
|---|---|---|
JUDGE_API_BASE |
— | e.g. https://your-endpoint.example/v1. Required. |
JUDGE_API_KEY |
— | Required. |
JUDGE_MODEL |
— | The judge model served there. Required. |
JUDGE_API_URL |
derived | Full chat-completions URL, for a service that does not use the /v1 layout. |
JUDGE_TEMPERATURE |
0.1 |
|
JUDGE_MAX_CONCURRENCY |
4 |
Per reward call. In training the aggregate is this times the trainer's reward-side limit. |
TARGET_MODEL_NAME |
placeholder | The model under test, as shown to the judge. |
TARGET_IDENTITY_NAMES |
empty | Comma-separated other names it may correctly use for itself. |
REQUIRE_THINK_CLOSE |
0 |
1 for a thinking checkpoint. |
Missing any of the three required values fails immediately rather than returning 0 for every sample.
2.8 Subcategories
M in a code marks the multi-turn counterpart of a single-turn subcategory.
| Ability | Subcategory | n | Turns | Subcategory | n | Turns | |
|---|---|---|---|---|---|---|---|
| DSLP 900 | DSLP-VCC-ICC | 200 | single | DSLP-VCC-MDCC | 50 | multi | |
| DSLP-VDMC-SVMR | 200 | single | DSLP-VDMC-MVMS | 50 | multi | ||
| DSLP-VTT-UCR | 200 | single | DSLP-VTT-UPC | 200 | single | ||
| MEA 900 | MEA-RC-AR | 300 | single | MEA-RC-MRR | 50 | multi | |
| MEA-RC-DR | 300 | single | MEA-SDI-MIST | 50 | multi | ||
| MEA-SDI-SPSD | 200 | single | |||||
| MSA 450 | MSA-IC-STIC | 200 | single | MSA-IC-MTIC | 50 | multi | |
| MSA-PLA † | 200 | single | |||||
| AH 400 | AH-AOR | 200 | single | AH-FGC | 200 | single | |
| ER 150 | ER | 150 | single |
† MSA-PLA carries two rubrics; see MSA-PLA: two rubrics.
Appendix D.5 of the paper defines each subcategory.
2.9 How the data was made
The dialogues were synthesized by OmniVChat-Studio, a multi-agent data engine: a Director handles text input and output, a Renderer turns an accepted prompt into a synchronized audio-visual clip, a Reviewer captions the result and writes a quality report, and a deterministic Validator checks scripts against the configured rules.

OmniVChat-Studio, with the single-turn subsystem (left) and the multi-turn subsystem (right). Each module carries the colour of the agent responsible for it.
2.10 Studio skills
Two English-language skills are available in skills/:
| Skill | Use |
|---|---|
| omnivchat-studio-build | Create, maintain, or refactor the Studio framework, its four agent roles, state transitions, media interfaces, and recovery behavior. |
| omnivchat-studio-add-subcategory | Add or extend a single-turn or multi-turn subcategory, including capability boundaries, diversity configuration, templates, rubrics, and validation. |
Each folder includes a SKILL.md entry point, supporting references, and agent metadata.
Load the appropriate skill in your coding agent with the target Studio implementation as
its workspace. The build skill also covers creating a new implementation. These are
development instructions; model and media backends are supplied by the operator.
3. OmniVChat-RL
The reward recipe used to train on synthesized dialogues, in reward/. Equation 4 of the
paper:
R(y) = r(y) + λ_fmt · f(y) + λ_eff · e(y) + λ_sty · s(y)
r, f, e, s score correctness, format, efficiency and style. The correctness term is
exactly the r(y) the benchmark reports. Template defaults are λ_fmt = 0.5, λ_eff = 0,
λ_sty = 0, so the training reward's range is [0, 1.5].
| Term | ||
|---|---|---|
r |
correctness | the tiered rubric, identical to the benchmark's score |
f |
format | binary: whether the completion's structure is valid (reasoning tags placed correctly, non-empty reply) |
e |
efficiency | reply-length efficiency, normalised within a GSPO rollout group. It does not exist for a single sample and cannot be computed offline. Off by default |
s |
style | style adherence, one extra judge call per sample. Off by default |
reward/reward.py is the reward — one file, framework-independent, and the same file
eval/score.py imports. reward/trainer_adapter.py is optional plumbing for wiring it into a
GSPO trainer (trainer_adapter.omni_basic_reward); it also implements e, the one term that
needs a whole rollout group, and defines no part of the metric. Run
python3 reward/trainer_adapter.py for the self-check. See reward/README.md.
3.1 Why the two scores are not comparable
- Scale. A training reward mean is not on the same scale as a benchmark score. Set
FORMAT_WEIGHT=0and the reward reduces exactly tor(y). - Judge failures. Evaluation excludes a sample whose judge call never succeeded from the mean. Training substitutes a running mean of the samples already scored. Scoring an API failure as 0 during training would punish a completion for something it did not do.
3.2 Training curves

OmniVChat-RL over 1,000 iterations. Top: the unweighted reward terms of Equation 4, raw reply efficiency, and mean reply length. Bottom: the gated rubric of Equation 1 every twentieth step, on the development set, on OmniVChat-Bench, and on OmniVChat-Bench-Human. The dashed line marks the selected step 940. Scores across the three corpora are not directly comparable.
Two things in that figure are worth reading together with the weights above. format sits at
1.000 for the whole run, so it contributes no within-group learning signal — which is the
behaviour the reward code's own comments predict. And mean reply length falls from about 100
words to about 35 while the rubric term keeps rising: raw efficiency, rubric credit per
thousand words, goes from 3.37 to 17.18. Shorter replies scoring better is the intended
effect of the efficiency term, not a side effect.
Judge configuration for both paths, via environment: JUDGE_API_BASE, JUDGE_API_KEY,
JUDGE_MODEL, TARGET_MODEL_NAME, TARGET_IDENTITY_NAMES. See
Configuring the judge.
4. Notes
- The dialogues are synthesized, not recorded. The paper also reports on a human-recorded probe, OmniVChat-Bench-Human (360 instances), which is not part of this repository.
- Rubrics were authored in Chinese and translated to English; the judge prompts and all code comments are in English.
- github.com/HarlandZZC/OmniVChat mirrors this
repository without
data/andmedia/, for reading the code without the 28 GB download.
5. License and Intended Use
This dataset is licensed under the
Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.
The full license text is provided in LICENSE.
This dataset is an evaluation benchmark released under a restricted license and may be used solely for academic research and model evaluation. Any form of commercial use is strictly prohibited. Images contained in the dataset may depict identifiable natural persons; copyright and all related rights in those images remain with their respective rights holders. Release of this dataset does not grant permission to relicense, commercially exploit, or create derivative works from any included image. Users must use the dataset within the scope of the foregoing license and permissions, must not use the dataset or any included image in commercial products, services, or other profit-making activities, and are solely responsible for complying with all applicable laws and regulations concerning personal information protection and portrait rights.
6. Acknowledgements
This work was supported by Alibaba Research Intern Program. We would like to thank the Qwen Team at Alibaba Token Hub (ATH), Alibaba Group, for providing the computational resources and foundation models (Qwen) used in this research.
7. Citation
@misc{he2026omnivchat,
title = {OmniVChat: Synthesizing, Benchmarking, and Training for Native Audio-Visual Dialogue},
author = {Haolin He and Yunfei Chu and Qi Chen and Wen Huang and Yuan Feng and Muzhi Zhu and Zheqi Dai and Haoning Xu and Dongchao Yang and Chunyat Wu and Zining Liang and Zhengxi Liu and Xiquan Li and Xie Chen and Xize Cheng and Qize Yang and Jin Xu and Qiuqiang Kong},
year = {2026},
eprint = {2609.21465},
archivePrefix = {arXiv},
primaryClass = {eess.AS},
url = {https://arxiv.org/abs/2609.21465}
}
- Downloads last month
- 2,556







