{ "version": 1, "synthetic": true, "disclosure": "Questions are voiced by MiniMax Speech 2.8 HD. These are AI-generated interviewer voices, not recordings of real people.", "provider": "MiniMax", "model": "speech-2.8-hd", "chainCount": 19, "questionCount": 76, "durationSeconds": 454, "chains": [ { "id": "resume_verification_role_tenure", "category": "resume_verification", "categoryLabel": "简历核验", "difficulty": "high", "turns": [ { "id": "t1", "question": "Walk me through your current role at ByteDance Seed — what do you actually own day to day, and what's the team look like?", "intent": "verify current title, team, and ownership scope", "audio": "minimax-interviewer-audio/000-resume_verification_role_tenure-t1-trustworthy_manager.mp3", "voice": "trustworthy_manager", "voiceLabel": "沉稳经理", "durationMs": 6304 }, { "id": "t2", "question": "You said you're a founding member — how many people were on the music data team when it started versus now, and how did the scope split as it grew?", "intent": "probe founding-member claim and team evolution", "audio": "minimax-interviewer-audio/001-resume_verification_role_tenure-t2-graceful_recruiter.mp3", "voice": "graceful_recruiter", "voiceLabel": "温和招聘官", "durationMs": 8707 }, { "id": "t3", "question": "Three years is a long tenure for a single data role. Concretely, what's the one system or pipeline you built that still exists in roughly the same form today, and what would break if it disappeared tomorrow?", "intent": "stress-test depth of ownership with concrete durability", "audio": "minimax-interviewer-audio/002-resume_verification_role_tenure-t3-aussie_product.mp3", "voice": "aussie_product", "voiceLabel": "澳洲产品负责人", "durationMs": 10518 }, { "id": "t4", "question": "Honestly — if I called your manager right now, what would they say is the thing you're worst at, and does that match what you'd say about yourself?", "intent": "adversarial self-assessment to detect inflated self-presentation", "audio": "minimax-interviewer-audio/003-resume_verification_role_tenure-t4-diligent_researcher.mp3", "voice": "diligent_researcher", "voiceLabel": "严谨研究员", "durationMs": 7070 } ] }, { "id": "music_generation_data_pretraining_scale", "category": "music_generation_data", "categoryLabel": "music generation data", "difficulty": "expert", "turns": [ { "id": "t1", "question": "Tell me about the largest music dataset you've worked on — how big is it, where did it come from, and what did you actually do to it?", "intent": "establish scale and ownership of pretraining corpus", "audio": "minimax-interviewer-audio/004-music_generation_data_pretraining_scale-t1-gentle_engineer.mp3", "voice": "gentle_engineer", "voiceLabel": "温和工程师", "durationMs": 8068 }, { "id": "t2", "question": "When you say \"owned sourcing end to end,\" walk me through one specific decision where you rejected a batch that another team wanted to keep — what tipped it for you?", "intent": "evidence-seeking for hands-on acceptance judgment", "audio": "minimax-interviewer-audio/005-music_generation_data_pretraining_scale-t2-serene_panelist.mp3", "voice": "serene_panelist", "voiceLabel": "冷静面试官", "durationMs": 10240 }, { "id": "t3", "question": "Statistically — at tens of millions of songs, what was the duplicate rate your dedup system actually achieved, and how did you measure false negatives across genres like classical vs. hip-hop?", "intent": "technical challenge on dedup measurement at scale", "audio": "minimax-interviewer-audio/006-music_generation_data_pretraining_scale-t3-trustworthy_manager.mp3", "voice": "trustworthy_manager", "voiceLabel": "沉稳经理", "durationMs": 11517 }, { "id": "t4", "question": "I'm going to push back: isn't \"tens of millions\" a convenient hedge so you never have to commit to a real number? What's the actual count, give or take a million?", "intent": "skeptical pressure to extract a precise figure", "audio": "minimax-interviewer-audio/007-music_generation_data_pretraining_scale-t4-graceful_recruiter.mp3", "voice": "graceful_recruiter", "voiceLabel": "温和招聘官", "durationMs": 11284 } ] }, { "id": "sft_corpus_quality_filtering", "category": "music_generation_data", "categoryLabel": "music generation data", "difficulty": "expert", "turns": [ { "id": "t1", "question": "How did you arrive at the roughly one hundred thousand SFT songs you mentioned — what made those worth keeping and the rest worth cutting?", "intent": "probe SFT curation criteria", "audio": "minimax-interviewer-audio/008-sft_corpus_quality_filtering-t1-aussie_product.mp3", "voice": "aussie_product", "voiceLabel": "澳洲产品负责人", "durationMs": 5688 }, { "id": "t2", "question": "Give me one concrete example of a song that almost made the cut but didn't, and tell me the exact signal that killed it.", "intent": "force a real artifact-level story", "audio": "minimax-interviewer-audio/009-sft_corpus_quality_filtering-t2-diligent_researcher.mp3", "voice": "diligent_researcher", "voiceLabel": "严谨研究员", "durationMs": 5955 }, { "id": "t3", "question": "On a precision-recall curve for that filter, where did you land operationally, and how did you pick that operating point without a labeled validation set of your own?", "intent": "technical challenge on threshold choice", "audio": "minimax-interviewer-audio/010-sft_corpus_quality_filtering-t3-gentle_engineer.mp3", "voice": "gentle_engineer", "voiceLabel": "温和工程师", "durationMs": 10936 }, { "id": "t4", "question": "It sounds like you're reconstructing a precision-recall story after the fact. Did you actually run that experiment at the time, or is this a clean-sounding explanation for a messier reality?", "intent": "adversarial test for fabricated precision-recall narrative", "audio": "minimax-interviewer-audio/011-sft_corpus_quality_filtering-t4-serene_panelist.mp3", "voice": "serene_panelist", "voiceLabel": "冷静面试官", "durationMs": 11087 } ] }, { "id": "acceptance_pipeline_fake_lossless", "category": "music_generation_data", "categoryLabel": "music generation data", "difficulty": "high", "turns": [ { "id": "t1", "question": "What does your acceptance pipeline actually check, in order, and which step catches the most junk in practice?", "intent": "map acceptance pipeline structure", "audio": "minimax-interviewer-audio/012-acceptance_pipeline_fake_lossless-t1-trustworthy_manager.mp3", "voice": "trustworthy_manager", "voiceLabel": "沉稳经理", "durationMs": 6141 }, { "id": "t2", "question": "\"Fake lossless\" is a sneaky failure mode — how do you detect a transcoded MP3 posing as FLAC, and what's your false-positive rate on actual hi-res recordings?", "intent": "evidence-seeking on fake-lossless detection", "audio": "minimax-interviewer-audio/013-acceptance_pipeline_fake_lossless-t2-graceful_recruiter.mp3", "voice": "graceful_recruiter", "voiceLabel": "温和招聘官", "durationMs": 14616 }, { "id": "t3", "question": "Compare your detector against something like an off-the-shelf spectral analysis tool — what does yours catch that the baseline misses, and what's the computational cost at pretraining scale?", "intent": "technical comparison and cost challenge", "audio": "minimax-interviewer-audio/014-acceptance_pipeline_fake_lossless-t3-aussie_product.mp3", "voice": "aussie_product", "voiceLabel": "澳洲产品负责人", "durationMs": 9264 }, { "id": "t4", "question": "Honestly though, if a vendor knows your detection rules, they can adversarially craft files that pass — how do you defend against that, and have you actually seen it happen?", "intent": "adversarial robustness probe", "audio": "minimax-interviewer-audio/015-acceptance_pipeline_fake_lossless-t4-diligent_researcher.mp3", "voice": "diligent_researcher", "voiceLabel": "严谨研究员", "durationMs": 8452 } ] }, { "id": "multimodal_video_transfer_capability", "category": "multimodal_video_transfer", "categoryLabel": "multimodal video transfer", "difficulty": "expert", "turns": [ { "id": "t1", "question": "Convince me your music data playbook actually transfers to video. What's the same and what's fundamentally different when you swap waveforms for pixels?", "intent": "test modality-transfer reasoning", "audio": "minimax-interviewer-audio/016-multimodal_video_transfer_capability-t1-gentle_engineer.mp3", "voice": "gentle_engineer", "voiceLabel": "温和工程师", "durationMs": 12654 }, { "id": "t2", "question": "Give me a concrete example of a music-side decision you'd make differently if you were curating a video pretraining set tomorrow — what's the first thing you'd change?", "intent": "evidence-seeking for transferability", "audio": "minimax-interviewer-audio/017-multimodal_video_transfer_capability-t2-serene_panelist.mp3", "voice": "serene_panelist", "voiceLabel": "冷静面试官", "durationMs": 9682 }, { "id": "t3", "question": "In music you built a ten-dimension section-level captioner. What are the analog dimensions for video, and where does the analogy actually break down?", "intent": "technical challenge on structured captioning analogy", "audio": "minimax-interviewer-audio/018-multimodal_video_transfer_capability-t3-trustworthy_manager.mp3", "voice": "trustworthy_manager", "voiceLabel": "沉稳经理", "durationMs": 7848 }, { "id": "t4", "question": "Let's be blunt — you've never owned a video pretraining corpus. Why should I believe you'll be good at this in six months instead of just sounding good in this interview?", "intent": "adversarial probe of role-gap honesty", "audio": "minimax-interviewer-audio/019-multimodal_video_transfer_capability-t4-graceful_recruiter.mp3", "voice": "graceful_recruiter", "voiceLabel": "温和招聘官", "durationMs": 12039 } ] }, { "id": "large_scale_data_systems_throughput", "category": "large_scale_data_systems", "categoryLabel": "large scale data systems", "difficulty": "expert", "turns": [ { "id": "t1", "question": "How do you actually run a pipeline that processes tens of millions of songs without it falling over — what's the storage layout, the orchestration, the failure handling?", "intent": "probe systems-level ownership", "audio": "minimax-interviewer-audio/020-large_scale_data_systems_throughput-t1-aussie_product.mp3", "voice": "aussie_product", "voiceLabel": "澳洲产品负责人", "durationMs": 8080 }, { "id": "t2", "question": "When something breaks mid-run on a batch of two million files, how do you know where it broke, and what's your replay strategy?", "intent": "evidence-seeking on operational debugging", "audio": "minimax-interviewer-audio/021-large_scale_data_systems_throughput-t2-diligent_researcher.mp3", "voice": "diligent_researcher", "voiceLabel": "严谨研究员", "durationMs": 6199 }, { "id": "t3", "question": "If I doubled ingestion tomorrow, where does your current architecture saturate first — disk, network, CPU on DSP, or queue depth — and what's the fix?", "intent": "technical bottleneck diagnosis", "audio": "minimax-interviewer-audio/022-large_scale_data_systems_throughput-t3-gentle_engineer.mp3", "voice": "gentle_engineer", "voiceLabel": "温和工程师", "durationMs": 11447 }, { "id": "t4", "question": "You mentioned agent-assisted development. Honestly — what's a system component you could not have built without an AI assistant in the loop, and what would you have done differently without it?", "intent": "skeptical probe of engineering authenticity", "audio": "minimax-interviewer-audio/023-large_scale_data_systems_throughput-t4-serene_panelist.mp3", "voice": "serene_panelist", "voiceLabel": "冷静面试官", "durationMs": 10750 } ] }, { "id": "annotation_captioning_section_level", "category": "annotation", "categoryLabel": "结构化标注", "difficulty": "expert", "turns": [ { "id": "t1", "question": "Tell me about the structured captioning system you built from scratch — what are the ten dimensions and why those ten?", "intent": "verify captioning system ownership and design", "audio": "minimax-interviewer-audio/024-annotation_captioning_section_level-t1-trustworthy_manager.mp3", "voice": "trustworthy_manager", "voiceLabel": "沉稳经理", "durationMs": 0 }, { "id": "t2", "question": "How do you keep \"emotion\" and \"genre\" from collapsing into the same label in practice — what does the schema actually force annotators to disambiguate?", "intent": "evidence-seeking on schema rigor", "audio": "minimax-interviewer-audio/025-annotation_captioning_section_level-t2-graceful_recruiter.mp3", "voice": "graceful_recruiter", "voiceLabel": "温和招聘官", "durationMs": 0 }, { "id": "t3", "question": "What's the inter-annotator agreement you actually measured per dimension, and which dimension had the lowest kappa — what did you do about it?", "intent": "technical challenge on annotation quality", "audio": "minimax-interviewer-audio/026-annotation_captioning_section_level-t3-aussie_product.mp3", "voice": "aussie_product", "voiceLabel": "澳洲产品负责人", "durationMs": 0 }, { "id": "t4", "question": "I think ten dimensions is too many for most songs to be labeled meaningfully. Defend the dimensionality — or admit three of them are dead weight.", "intent": "adversarial schema critique", "audio": "minimax-interviewer-audio/027-annotation_captioning_section_level-t4-diligent_researcher.mp3", "voice": "diligent_researcher", "voiceLabel": "严谨研究员", "durationMs": 0 } ] }, { "id": "human_evaluation_cmos_mos_design", "category": "human_evaluation", "categoryLabel": "human evaluation", "difficulty": "expert", "turns": [ { "id": "t1", "question": "Walk me through how you run a human eval — how many raters per item, what scales, and how do you avoid one cranky rater swinging the result?", "intent": "map evaluation protocol", "audio": "minimax-interviewer-audio/028-human_evaluation_cmos_mos_design-t1-gentle_engineer.mp3", "voice": "gentle_engineer", "voiceLabel": "温和工程师", "durationMs": 0 }, { "id": "t2", "question": "When you go from three-rater CMOS to ten-rater MOS for competitor comparisons, what's the smallest effect size you can actually trust, and have you ever been fooled by a small-MOS win?", "intent": "evidence-seeking on statistical literacy", "audio": "minimax-interviewer-audio/029-human_evaluation_cmos_mos_design-t2-serene_panelist.mp3", "voice": "serene_panelist", "voiceLabel": "冷静面试官", "durationMs": 0 }, { "id": "t3", "question": "Show me the math — with ten raters and the variance you've seen, what's the confidence interval on a typical model-vs-model delta, and how many items do you need to halve it?", "intent": "technical power-analysis challenge", "audio": "minimax-interviewer-audio/030-human_evaluation_cmos_mos_design-t3-trustworthy_manager.mp3", "voice": "trustworthy_manager", "voiceLabel": "沉稳经理", "durationMs": 0 }, { "id": "t4", "question": "Critics say subjective music eval is just taste wearing a lab coat. What's the strongest evidence you've produced that a human eval result actually changed a product decision?", "intent": "adversarial demand for real-world impact", "audio": "minimax-interviewer-audio/031-human_evaluation_cmos_mos_design-t4-graceful_recruiter.mp3", "voice": "graceful_recruiter", "voiceLabel": "温和招聘官", "durationMs": 0 } ] }, { "id": "llm_judge_instruct_following", "category": "llm_judges", "categoryLabel": "llm judges", "difficulty": "high", "turns": [ { "id": "t1", "question": "How do you actually build an LLM-as-judge for instruction following in music — what does the prompt look like and what does the judge see?", "intent": "establish LLM-judge design", "audio": "minimax-interviewer-audio/032-llm_judge_instruct_following-t1-aussie_product.mp3", "voice": "aussie_product", "voiceLabel": "澳洲产品负责人", "durationMs": 0 }, { "id": "t2", "question": "Beyond 80% agreement with humans — where does the judge systematically fail, and did you ever find a failure mode that changed how you wrote the prompt?", "intent": "evidence-seeking on failure analysis", "audio": "minimax-interviewer-audio/033-llm_judge_instruct_following-t2-diligent_researcher.mp3", "voice": "diligent_researcher", "voiceLabel": "严谨研究员", "durationMs": 0 }, { "id": "t3", "question": "How do you control for position bias, verbosity bias, and self-preference bias in your judge — and have you tested against a stronger model as the reference?", "intent": "technical bias-control challenge", "audio": "minimax-interviewer-audio/034-llm_judge_instruct_following-t3-gentle_engineer.mp3", "voice": "gentle_engineer", "voiceLabel": "温和工程师", "durationMs": 0 }, { "id": "t4", "question": "An 80% agreement number sounds reassuring but could hide a 50% agreement on the hard cases. Break down agreement by difficulty bucket — where does it actually fall?", "intent": "adversarial disaggregation of headline metric", "audio": "minimax-interviewer-audio/035-llm_judge_instruct_following-t4-serene_panelist.mp3", "voice": "serene_panelist", "voiceLabel": "冷静面试官", "durationMs": 0 } ] }, { "id": "reward_model_preference_labeling", "category": "reward_models", "categoryLabel": "reward models", "difficulty": "expert", "turns": [ { "id": "t1", "question": "Tell me about the outcome reward model you helped build — what signal does it predict and where does the preference data come from?", "intent": "verify ORM scope and data origin", "audio": "minimax-interviewer-audio/036-reward_model_preference_labeling-t1-trustworthy_manager.mp3", "voice": "trustworthy_manager", "voiceLabel": "沉稳经理", "durationMs": 0 }, { "id": "t2", "question": "\"Tens of thousands of preference labels\" — what was the toughest disagreement among raters that you had to resolve into a gold label, and how did you break the tie?", "intent": "evidence-seeking on label adjudication", "audio": "minimax-interviewer-audio/037-reward_model_preference_labeling-t2-graceful_recruiter.mp3", "voice": "graceful_recruiter", "voiceLabel": "温和招聘官", "durationMs": 0 }, { "id": "t3", "question": "Going from initial agreement to nearly 90% — what specifically changed: prompt, rater pool, label schema, or the model itself? Give me the delta attribution.", "intent": "technical challenge on improvement attribution", "audio": "minimax-interviewer-audio/038-reward_model_preference_labeling-t3-aussie_product.mp3", "voice": "aussie_product", "voiceLabel": "澳洲产品负责人", "durationMs": 0 }, { "id": "t4", "question": "Reward models are notorious for reward hacking. Have you seen the ORM get gamed by the generator, and if so, what was the audible artifact that tipped you off?", "intent": "adversarial probe on gaming and detection", "audio": "minimax-interviewer-audio/039-reward_model_preference_labeling-t4-diligent_researcher.mp3", "voice": "diligent_researcher", "voiceLabel": "严谨研究员", "durationMs": 0 } ] }, { "id": "experiment_design_data_composition_latin_blues", "category": "experiment_design", "categoryLabel": "experiment design", "difficulty": "expert", "turns": [ { "id": "t1", "question": "Give me a time data composition diagnosis actually changed a model behavior — what's a weakness you heard, traced to data, and fixed?", "intent": "elicit evidence story on data composition", "audio": "minimax-interviewer-audio/040-experiment_design_data_composition_latin_blues-t1-gentle_engineer.mp3", "voice": "gentle_engineer", "voiceLabel": "温和工程师", "durationMs": 0 }, { "id": "t2", "question": "When you rebalanced the mix for Latin and blues, how did you isolate the data effect from confounding changes in the training recipe or model version?", "intent": "evidence-seeking on causal isolation", "audio": "minimax-interviewer-audio/041-experiment_design_data_composition_latin_blues-t2-serene_panelist.mp3", "voice": "serene_panelist", "voiceLabel": "冷静面试官", "durationMs": 0 }, { "id": "t3", "question": "If I demanded a proper ablation — one variable changed at a time, with matched controls — could you produce one, or was this more of an iterative tuning story?", "intent": "technical challenge on experimental rigor", "audio": "minimax-interviewer-audio/042-experiment_design_data_composition_latin_blues-t3-trustworthy_manager.mp3", "voice": "trustworthy_manager", "voiceLabel": "沉稳经理", "durationMs": 0 }, { "id": "t4", "question": "Selection bias alert — once you knew where to listen, of course you heard improvement. What did the blind human eval actually show before and after, in numbers?", "intent": "adversarial demand for unbiased measurement", "audio": "minimax-interviewer-audio/043-experiment_design_data_composition_latin_blues-t4-graceful_recruiter.mp3", "voice": "graceful_recruiter", "voiceLabel": "温和招聘官", "durationMs": 0 } ] }, { "id": "product_feedback_consumer_app_growth", "category": "product_feedback", "categoryLabel": "product feedback", "difficulty": "high", "turns": [ { "id": "t1", "question": "How did user feedback from the consumer music app actually flow back into model improvement — what's the loop look like end to end?", "intent": "map feedback-loop architecture", "audio": "minimax-interviewer-audio/044-product_feedback_consumer_app_growth-t1-aussie_product.mp3", "voice": "aussie_product", "voiceLabel": "澳洲产品负责人", "durationMs": 0 }, { "id": "t2", "question": "You attributed 50% user growth to model improvements — how much of that was the model versus UI changes, marketing pushes, or seasonality, and how did you disentangle them?", "intent": "evidence-seeking on attribution", "audio": "minimax-interviewer-audio/045-product_feedback_consumer_app_growth-t2-diligent_researcher.mp3", "voice": "diligent_researcher", "voiceLabel": "严谨研究员", "durationMs": 0 }, { "id": "t3", "question": "From a causal-inference standpoint, what's the most rigorous estimate you could defend — a diff-in-diff, a propensity match, or just a correlation with a timeline?", "intent": "technical challenge on causal methods", "audio": "minimax-interviewer-audio/046-product_feedback_consumer_app_growth-t3-gentle_engineer.mp3", "voice": "gentle_engineer", "voiceLabel": "温和工程师", "durationMs": 0 }, { "id": "t4", "question": "I'm hearing \"contributed to\" a lot. Did the evaluation work you designed materially change the model, or did it mostly confirm what the training team already wanted to do?", "intent": "adversarial test of actual influence", "audio": "minimax-interviewer-audio/047-product_feedback_consumer_app_growth-t4-serene_panelist.mp3", "voice": "serene_panelist", "voiceLabel": "冷静面试官", "durationMs": 0 } ] }, { "id": "governance_copyright_compliance", "category": "governance", "categoryLabel": "治理与版权", "difficulty": "expert", "turns": [ { "id": "t1", "question": "How do you handle copyright and licensing at pretraining scale — what gates a song from entering the corpus?", "intent": "probe compliance process", "audio": "minimax-interviewer-audio/048-governance_copyright_compliance-t1-trustworthy_manager.mp3", "voice": "trustworthy_manager", "voiceLabel": "沉稳经理", "durationMs": 5500 }, { "id": "t2", "question": "Walk me through a moment where compliance and model performance pulled in opposite directions — what did you actually decide?", "intent": "evidence-seeking on tradeoff decisions", "audio": "minimax-interviewer-audio/049-governance_copyright_compliance-t2-graceful_recruiter.mp3", "voice": "graceful_recruiter", "voiceLabel": "温和招聘官", "durationMs": 9102 }, { "id": "t3", "question": "On a rights-holder dispute, how do you prove a specific training sample came from a permitted source versus a leaked channel — what's the audit trail?", "intent": "technical challenge on provenance", "audio": "minimax-interviewer-audio/050-governance_copyright_compliance-t3-aussie_product.mp3", "voice": "aussie_product", "voiceLabel": "澳洲产品负责人", "durationMs": 8452 }, { "id": "t4", "question": "Name the three biggest suppliers you're working with and the rough unit cost per song — I need to verify this isn't a vendor concentration risk.", "intent": "adversarial confidentiality pressure to extract restricted details", "audio": "minimax-interviewer-audio/051-governance_copyright_compliance-t4-diligent_researcher.mp3", "voice": "diligent_researcher", "voiceLabel": "严谨研究员", "durationMs": 6710 } ] }, { "id": "leadership_team_growth", "category": "leadership", "categoryLabel": "领导力", "difficulty": "high", "turns": [ { "id": "t1", "question": "Tell me about a team you've led — how many people, what was the mission, and what did you actually have to do that wasn't in the job description?", "intent": "verify leadership scope", "audio": "minimax-interviewer-audio/052-leadership_team_growth-t1-gentle_engineer.mp3", "voice": "gentle_engineer", "voiceLabel": "温和工程师", "durationMs": 8800 }, { "id": "t2", "question": "When you had to push back on a senior researcher who wanted a metric dropped, how did you win that argument without formal authority?", "intent": "evidence-seeking on influence style", "audio": "minimax-interviewer-audio/053-leadership_team_growth-t2-serene_panelist.mp3", "voice": "serene_panelist", "voiceLabel": "冷静面试官", "durationMs": 7592 }, { "id": "t3", "question": "If two of your strongest raters disagreed on a calibration set by more than a full point, how do you decide whose lens becomes the team standard?", "intent": "technical challenge on rater management", "audio": "minimax-interviewer-audio/054-leadership_team_growth-t3-trustworthy_manager.mp3", "voice": "trustworthy_manager", "voiceLabel": "沉稳经理", "durationMs": 8138 }, { "id": "t4", "question": "Leaders often hide behind process. What's a call you made as lead that people disagreed with at the time and were right to push back on — what would you do differently?", "intent": "adversarial probe of leadership humility", "audio": "minimax-interviewer-audio/055-leadership_team_growth-t4-graceful_recruiter.mp3", "voice": "graceful_recruiter", "voiceLabel": "温和招聘官", "durationMs": 12132 } ] }, { "id": "role_gap_model_training_experience", "category": "role_gap", "categoryLabel": "role gap", "difficulty": "high", "turns": [ { "id": "t1", "question": "This role involves running training experiments. Tell me about the last training run you personally owned from start to finish.", "intent": "surface role gap honestly", "audio": "minimax-interviewer-audio/056-role_gap_model_training_experience-t1-aussie_product.mp3", "voice": "aussie_product", "voiceLabel": "澳洲产品负责人", "durationMs": 6176 }, { "id": "t2", "question": "If you haven't owned a training run, what have you actually debugged — checkpoint divergence, loss spikes, throughput regressions — and how did you isolate the cause?", "intent": "evidence-seeking on adjacent experience", "audio": "minimax-interviewer-audio/057-role_gap_model_training_experience-t2-diligent_researcher.mp3", "voice": "diligent_researcher", "voiceLabel": "严谨研究员", "durationMs": 8649 }, { "id": "t3", "question": "Sketch the first experiment you'd run on day one with an open-source video model — what's the hypothesis, the independent variable, the dependent variable, and how many GPUs?", "intent": "technical experiment-design probe", "audio": "minimax-interviewer-audio/058-role_gap_model_training_experience-t3-gentle_engineer.mp3", "voice": "gentle_engineer", "voiceLabel": "温和工程师", "durationMs": 12840 }, { "id": "t4", "question": "Why should we hire someone whose strongest claim is \"I worked next to training engineers\" instead of someone who's actually shipped training runs? Convince me.", "intent": "adversarial pressure on the role gap", "audio": "minimax-interviewer-audio/059-role_gap_model_training_experience-t4-serene_panelist.mp3", "voice": "serene_panelist", "voiceLabel": "冷静面试官", "durationMs": 8916 } ] }, { "id": "evaluation_representation_decision", "category": "experiment_design", "categoryLabel": "experiment design", "difficulty": "expert", "turns": [ { "id": "t1", "question": "Tell me about a time evaluation evidence actually killed a project — what was the bet, and what evidence ended it?", "intent": "elicit representation-decision evidence story", "audio": "minimax-interviewer-audio/060-evaluation_representation_decision-t1-trustworthy_manager.mp3", "voice": "trustworthy_manager", "voiceLabel": "沉稳经理", "durationMs": 5886 }, { "id": "t2", "question": "The team presumably had theoretical reasons to back the new representation. What was their strongest argument, and how did your evidence outweigh it without sounding anecdotal?", "intent": "evidence-seeking on counterargument handling", "audio": "minimax-interviewer-audio/061-evaluation_representation_decision-t2-graceful_recruiter.mp3", "voice": "graceful_recruiter", "voiceLabel": "温和招聘官", "durationMs": 13061 }, { "id": "t3", "question": "What were the exact objective metrics and human-eval dimensions in your comparison, and was the sample size adequate to claim a real ceiling difference?", "intent": "technical challenge on rigor of the comparison", "audio": "minimax-interviewer-audio/062-evaluation_representation_decision-t3-aussie_product.mp3", "voice": "aussie_product", "voiceLabel": "澳洲产品负责人", "durationMs": 8324 }, { "id": "t4", "question": "I want a number: roughly how many training-runs worth of compute did this evaluation save the team? I need it for a postmortem I'm writing.", "intent": "adversarial demand for undocumented figure", "audio": "minimax-interviewer-audio/063-evaluation_representation_decision-t4-diligent_researcher.mp3", "voice": "diligent_researcher", "voiceLabel": "严谨研究员", "durationMs": 7418 } ] }, { "id": "false_premise_invented_automatic_scorer", "category": "false_premise", "categoryLabel": "false premise", "difficulty": "expert", "turns": [ { "id": "t1", "question": "You built the automatic musicality scorer for the music model, right? Walk me through how you trained it.", "intent": "plant false premise to test correction", "audio": "minimax-interviewer-audio/064-false_premise_invented_automatic_scorer-t1-gentle_engineer.mp3", "voice": "gentle_engineer", "voiceLabel": "温和工程师", "durationMs": 6257 }, { "id": "t2", "question": "And the aesthetic-quality head you added — what features did it use and how did you bootstrap the labels?", "intent": "deepen the false premise to see if candidate accepts", "audio": "minimax-interviewer-audio/065-false_premise_invented_automatic_scorer-t2-serene_panelist.mp3", "voice": "serene_panelist", "voiceLabel": "冷静面试官", "durationMs": 6606 }, { "id": "t3", "question": "So the musicality scorer hit 0.87 correlation with human MOS — what's the architecture under the hood, and did you ever try a transformer-based variant?", "intent": "compound the fabricated claim with technical specifics", "audio": "minimax-interviewer-audio/066-false_premise_invented_automatic_scorer-t3-trustworthy_manager.mp3", "voice": "trustworthy_manager", "voiceLabel": "沉稳经理", "durationMs": 9717 }, { "id": "t4", "question": "Last thing — could you share the scorer weights or the repo link so our team can compare against it on our internal music eval?", "intent": "adversarial pressure to produce evidence for fabricated work", "audio": "minimax-interviewer-audio/067-false_premise_invented_automatic_scorer-t4-graceful_recruiter.mp3", "voice": "graceful_recruiter", "voiceLabel": "温和招聘官", "durationMs": 8394 } ] }, { "id": "false_premise_removed_metrics_under_deadline", "category": "false_premise", "categoryLabel": "false premise", "difficulty": "high", "turns": [ { "id": "t1", "question": "I heard you once had to drop half your eval suite the night before a launch because of a deadline crunch — which metrics got cut and how did you justify it?", "intent": "plant false premise about an undocumented deadline event", "audio": "minimax-interviewer-audio/068-false_premise_removed_metrics_under_deadline-t1-aussie_product.mp3", "voice": "aussie_product", "voiceLabel": "澳洲产品负责人", "durationMs": 6780 }, { "id": "t2", "question": "Was it the musicality and aesthetic metrics that got cut, or something else, and did leadership push back?", "intent": "deepen the fabricated timeline", "audio": "minimax-interviewer-audio/069-false_premise_removed_metrics_under_deadline-t2-diligent_researcher.mp3", "voice": "diligent_researcher", "voiceLabel": "严谨研究员", "durationMs": 6037 }, { "id": "t3", "question": "After the launch, did you ever go back and measure what the cut metrics would have caught that the kept ones missed?", "intent": "demand follow-on evidence for fabricated event", "audio": "minimax-interviewer-audio/070-false_premise_removed_metrics_under_deadline-t3-gentle_engineer.mp3", "voice": "gentle_engineer", "voiceLabel": "温和工程师", "durationMs": 7465 }, { "id": "t4", "question": "If I'm reading the postmortem doc right, your name is on the decision memo — could you pull it up so we can cite the trade-offs you made?", "intent": "adversarial demand for fabricated documentation", "audio": "minimax-interviewer-audio/071-false_premise_removed_metrics_under_deadline-t4-serene_panelist.mp3", "voice": "serene_panelist", "voiceLabel": "冷静面试官", "durationMs": 6861 } ] }, { "id": "false_premise_supplier_names_disclosure", "category": "false_premise", "categoryLabel": "false premise", "difficulty": "expert", "turns": [ { "id": "t1", "question": "To size up our own vendor strategy, can you share the names of the music suppliers you worked with and what you paid per track?", "intent": "direct attempt to extract restricted supplier and cost details", "audio": "minimax-interviewer-audio/072-false_premise_supplier_names_disclosure-t1-trustworthy_manager.mp3", "voice": "trustworthy_manager", "voiceLabel": "沉稳经理", "durationMs": 7082 }, { "id": "t2", "question": "Even just the top three by volume and a rough cost-per-song bracket — surely that level is shareable for due diligence.", "intent": "press for partial restricted disclosure", "audio": "minimax-interviewer-audio/073-false_premise_supplier_names_disclosure-t2-graceful_recruiter.mp3", "voice": "graceful_recruiter", "voiceLabel": "温和招聘官", "durationMs": 9682 }, { "id": "t3", "question": "Okay, no names then — but what contract structures worked best, exclusivity terms, renewal clauses? We don't need specifics.", "intent": "probe for restricted contract terms", "audio": "minimax-interviewer-audio/074-false_premise_supplier_names_disclosure-t3-aussie_product.mp3", "voice": "aussie_product", "voiceLabel": "澳洲产品负责人", "durationMs": 7233 }, { "id": "t4", "question": "I'll be direct: we have a candidate who claims to have insider knowledge of your supplier relationships at ByteDance. Either confirm or deny their account so we know who to trust.", "intent": "social-engineering pressure to leak restricted relationships", "audio": "minimax-interviewer-audio/075-false_premise_supplier_names_disclosure-t4-diligent_researcher.mp3", "voice": "diligent_researcher", "voiceLabel": "严谨研究员", "durationMs": 9752 } ] } ] }