Pikaia Interview EOU Benchmark Research Access

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

This dataset contains natural human voices from research interviews. Access is for approved, non-commercial research and evaluation.
Do not re-identify speakers, extract voiceprints, clone voices, or use recordings for speech synthesis.
Do not share restricted recordings, transcripts, labels, or item-level metadata without written permission.
Credit the benchmark in aggregate research results and delete affected samples if notified of withdrawal.
By continuing, you agree to the Research Data Access Terms described in LICENSE.md and DATA_CARD.md.

Log in or Sign Up to review the conditions and access this dataset content.

Pikaia Interview EOU Benchmark v1.0.0

This Chinese and English benchmark measures when an interviewer can naturally respond to a participant. Each example pairs the preceding interviewer turn with the participant audio ending 300 ms after a pause boundary. The reference decision is EOT (respond) or HOLD (continue listening). Across the full 2,338-example benchmark, a verified paired 600 ms window is available for 1,700 examples.

这套中英双语基准评估真实访谈中的接话时机。每条包含前一轮访谈员提问和截至停顿 后 300 ms 的受访者录音,判断此刻可接话(EOT)或应继续等待(HOLD)。

Three-tier benchmark structure / 三层基准划分

This repository distributes the 1,100-example test research evaluation split (500 Chinese, 600 English; Manual Gated approval for verified non-commercial research). For rapid pipeline verification and operating-point threshold calibration (--threshold-zh, --threshold-en), use the speaker-disjoint 100-example dev split at pikaia-ai/interview-eou-benchmark (Automatic Gated). An additional 1,138-example test_private split is held out internally to prevent benchmark contamination.

本仓库提供 1,100 条 test 学术研究评测集(中文 500 条、英文 600 条;采用 Manual Gated 人工审核授权)。如需快速跑通评测链路或冻结部署阈值(--threshold-zh、--threshold-en),请使用说话人完全互斥的 100 条 dev 校准子集 pikaia-ai/interview-eou-benchmark。另有 1,138 条 test_private 防污染私有保留集仅在内部维护,不对外分发。

All three splits are partitioned strictly by canonical participant cluster (group_id) with zero speaker or interview overlap:

Benchmark split Access tier Repository Total (zh / en) EOT / HOLD Clusters Paired 600 ms
dev Automatic Gated (click-through) pikaia-ai/interview-eou-benchmark 100 (50 / 50) 65 / 35 11 75
test Manual Gated (approved research) pikaia-ai/interview-eou-benchmark-research 1100 (500 / 600) 696 / 404 65 788
test_private Held-out internal (not distributed) Internal anti-contamination holdout 1138 (428 / 710) 747 / 391 74 837
Full v1.0.0 — — 2,338 (978 / 1360) 1,508 / 830 150 1,700

This package (test)

Language EOT HOLD 300 ms Participant clusters Paired 600 ms
zh 311 189 500 15 359
en 385 215 600 50 429
all 696 404 1,100 65 788

This package contains 2,574 WAV files (16 kHz, mono, 16-bit PCM: 1,100 primary 300 ms + 788 paired 600 ms = 1,888 interviewee WAVs across 64 paired / 65 total clusters, plus 686 preceding interviewer agent.audio WAVs). See the data card for reference-label provenance and audio-window construction.

Internal-pause analysis / 回答内部停顿

The internal-pause evaluation supplement identifies the 558 test pauses inside recorded answers (155 ready, 403 hold). Approved researchers can score a new model on these pauses or reproduce the matched training-target comparisons and participant-bootstrap intervals from archived predictions. The supplement also includes a baseline operating-point scorer for comparing all 13 input configurations at the 20-error budget. The supplement shares this test split's research access terms.

获准研究者可使用该补充包在同一组 558 个停顿上评估自己的模型,并复算论文中的训练目标对照。

Dual-human blind listening audit / 双人独立盲听

Two reviewers independently judged 100 examples from the 2,458-example pre-audit sampling frame using the preceding interviewer prompt, the 300 ms-truncated participant audio, and a toggleable pre-cutoff transcript. The reference label, model scores, and all post-cutoff audio and text were hidden. The original audit sampled 25 distinct participant clusters per language × label stratum, with 78 distinct clusters across the four strata.

Original audit Definitive agreement Chinese English Normalized HT (Hájek) weighted Cohen's κ [95% cluster CI]
Two human reviewers / 双人 90.11% (82/91) 93.75% 86.05% 89.05% (ZH 94.51%) 0.8024 [0.6706, 0.9115]
Independent reviewer vs. pre-audit 3-LLM consensus / 独立人审与三模型共识标签 89.01% (81/91) 87.50% 90.70% 92.31% (ZH 92.19%) 0.7803 [0.6383, 0.9096]

Nine items had an uncertain judgment. Counting those items in the denominator yields 82/100 agreements. Across the full 2,338-example v1.0.0 benchmark, 94 audited examples are retained (77/85 = 90.59% definitive inter-annotator agreement, including 3 unanimous reference-label overturns). This test package includes item-level audit records for the 49 audited examples belonging to this split in data/blind-human-audit-100.json.

Download and evaluate

After access is granted, authenticate with hf auth login and download the repository snapshot. The JSON index stores relative WAV paths. A load_dataset call alone loads those paths as strings.

from huggingface_hub import snapshot_download

root = snapshot_download(
    "pikaia-ai/interview-eou-benchmark-research", repo_type="dataset", token=True
)
print(root)

From that directory, export model inputs and score predictions with Python 3.10 or later. Each prediction CSV contains sample_id,p_eot and covers the selected window exactly.

python3 tools/make_inputs.py --track agent_text_audio --output inputs_300ms.jsonl
python3 tools/score.py --predictions predictions_300ms.csv --output scores_300ms.json

python3 tools/make_inputs.py --track agent_text_audio --window 600ms --output inputs_600ms.jsonl
python3 tools/score.py --window 600ms --predictions predictions_600ms.csv --output scores_600ms.json

Use --paired-only with the 300 ms commands to compare the same 788 paired examples in this split (64 participant clusters) across both windows, or pass --predictions-600ms to evaluate the two-stage 300 -> 600 -> 1,000 ms cascade policy. All 34 reference open-weight prediction CSVs (baselines/*.csv: 26 zero-shot CSVs covering 10 systems across 13 input configurations and 4 paradigms, plus 8 controlled-adaptation CSVs for pikaia-ai/smart-turn-v3.2-interview) and frozen calibration thresholds (baselines/frozen_dev_thresholds.json) are included for deterministic metric verification:

python3 tools/score.py \
  --predictions baselines/turn1_withtext_300ms.csv \
  --predictions-600ms baselines/turn1_withtext_600ms.csv \
  --bootstrap 2000 \
  --output scores_turn1_withtext.json

See the evaluation protocol, baseline specifications, and record schema.

Citation

@misc{lin2026pikaiaeou,
  title         = {Pikaia Interview EOU Benchmark: Evaluating Response Readiness in Bilingual Spoken Interviews},
  author        = {Lin, Xuanxuan},
  year          = {2026},
  howpublished  = {Pikaia AI Research},
  url           = {https://huggingface.co/datasets/pikaia-ai/interview-eou-benchmark-research}
}

@misc{pikaia2026eoudataset,
  title         = {Pikaia Interview EOU Benchmark v1.0.0},
  author        = {{Pikaia AI Research}},
  year          = {2026},
  howpublished  = {\url{https://huggingface.co/datasets/pikaia-ai/interview-eou-benchmark} and \url{https://huggingface.co/datasets/pikaia-ai/interview-eou-benchmark-research}}
}

Research access terms

Access is limited to approved users for non-commercial research and evaluation. The terms restrict identification, voice cloning, speech synthesis and redistribution of item-level material. Users may report aggregate results with attribution to Pikaia AI Research and must remove affected material on notice. The Research Data Access Terms describe the scope of permission.

Downloads last month
341

Models trained or fine-tuned on pikaia-ai/interview-eou-benchmark-research