Add dataset card, labels, SVO-AQA, and metadata

.gitattributes

@@ -58,3 +58,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+metadata/train.jsonl filter=lfs diff=lfs merge=lfs -text
+train/train_label.json filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,117 @@
+---
+pretty_name: DeepASMR-NSpeech
+version: 1.0.0
+license: cc-by-nc-4.0
+language:
+- en
+task_categories:
+- text-to-audio
+- audio-classification
+configs:
+- config_name: default
+  data_files:
+  - split: train
+    path: "parquet/train-*.parquet"
+  - split: test
+    path: "parquet/test-*.parquet"
+---
+
+# DeepASMR-NSpeech
+
+DeepASMR-NSpeech is a non-speech ASMR audio dataset with structured
+Subject-Verb-Object annotations and an SVO-AQA audio multiple-choice benchmark.
+
+Version 1.0.0 was created by Ziyi Yang, Leying Zhang, Chenda Li, and Yanmin
+Qian at Shanghai Jiao Tong University.
+
+## Dataset size
+
+| Metric | Train | Test | Total / cross-split |
+|---|---:|---:|---:|
+| Source videos | 213 | 297 | 6 shared |
+| Creators | 14 | 6 | 4 shared |
+| Annotated events | 12,327 | 1,163 | 13,490 |
+| Ten-second clips | 72,666 | 1,163 | 73,829; 0 overlap |
+| Duration | 199.4 h | 3.19 h | 202.6 h |
+
+## Contents
+
+```text
+DeepASMR-NSpeech-dataset/
+  parquet/train-*.parquet
+  parquet/test-*.parquet
+  metadata/train.jsonl
+  metadata/test.jsonl
+  train/train_label.json
+  test/test_label.json
+  SVO-AQA/test_subset/bank.json
+  SVO-AQA/test_subset/audio_dependent_subset.v1.json
+  vocab/closed_vocab_reference.md
+  release_manifest.json
+  parquet_manifest.json
+```
+
+The Parquet shards contain embedded audio bytes and the public fields
+`audio_id`, `audio`, `caption`, `verb`, `subject`, `object`, and `creator_id`.
+This follows the public DeepASMR release pattern and avoids publishing
+machine-specific source paths.
+
+```python
+from datasets import load_dataset
+
+dataset = load_dataset("yzyai/DeepASMR-NSpeech")
+sample = dataset["train"][0]
+print(sample["caption"], sample["audio"])
+```
+
+Clip-label files contain relative `wav`, `caption`, closed-set `verb`, open noun
+phrases `subject` and `object`, and anonymized `creator_id`. Raw creator timeline
+text, titles, channel names, API metadata, and private quality-control GT are not
+part of the public label files.
+
+## Creation
+
+The public code repository provides the complete collection and annotation
+path: timeline acquisition, normalization/translation, segmentation, keyframe
+extraction, multimodal annotation, vocabulary normalization, label export, and
+SVO-AQA construction. SVO-AQA starts from human base annotations and can be
+extended by adding rows following the documented schema.
+
+The repository also provides annotation-quality audit code. Its GT must be
+independently labeled by humans; the authors' private GT is not distributed.
+
+## Evaluation
+
+For SVO-AQA, send each question's audio and prompt to a model and compare its
+single-letter answer with `gold_label`. Report overall and per-task accuracy.
+The full bank contains 431 human-validated questions. The paper's main table
+uses the frozen 55-question audio-dependent subset distributed beside the bank;
+the GitHub repository provides the judge, subset-selection, and multi-run
+scoring code.
+
+For dataset-label quality, report strict metrics against manual GT separately
+from any alias- or LLM-assisted entity-resolution score. Processing success
+counts are operational statistics, not label-quality metrics.
+
+## License, limitations, and source-media rights
+
+The structured annotations, dataset metadata, and SVO-AQA benchmark annotations
+that the authors are authorized to license are released under
+[CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/). Structured
+labels may contain model-assisted annotation noise.
+
+The source media originates from third-party public videos. All copyright and
+associated rights in the original source videos, original audio, and
+creator-produced content remain with their respective original creators or
+other rights holders. CC BY-NC 4.0 does not transfer ownership of that content
+or grant rights that the dataset authors do not hold. Users must comply with
+applicable creator rights, platform terms, privacy requirements, and local law.
+
+## Links
+
+- Code: `https://github.com/bonnie-yzy/DeepASMR-NSpeech`
+- Dataset: `https://huggingface.co/datasets/yzyai/DeepASMR-NSpeech`
+- Paper: `TBD`
+
+Replace the paper URL after publication, and keep the paper and repository
+rights statements aligned.

SVO-AQA/test_subset/audio_dependent_subset.v1.json

@@ -0,0 +1,417 @@
+{
+  "meta": {
+    "version": "1.0.0",
+    "selection": "reference_correct_votes >= 2 of 3 and silence_correct_votes <= 1 of 3",
+    "selection_judge_models": [
+      "qwen3.5-omni-plus-2026-03-15"
+    ],
+    "scoring": "highest first-nonempty-token log-probability",
+    "generated_audio_preprocessing": "48 kHz, per-clip RMS matched to reference",
+    "n_questions": 55,
+    "n_clips": 50,
+    "n_by_task": {
+      "object_material_mc": 11,
+      "subject_material_mc": 3,
+      "verb_mc": 41
+    },
+    "reference_accuracy_by_task": {
+      "object_material_mc": 1.0,
+      "subject_material_mc": 0.888889,
+      "verb_mc": 1.0
+    },
+    "silence_accuracy_by_task": {
+      "object_material_mc": 0.0,
+      "subject_material_mc": 0.0,
+      "verb_mc": 0.03252
+    },
+    "reference_macro_accuracy": 0.962963,
+    "silence_macro_accuracy": 0.01084
+  },
+  "questions": [
+    {
+      "question_id": "om__fingers__tapping_____dglass-wood__inst0",
+      "audio_id": "2NnjZFbCxG0_757_988",
+      "task": "object_material_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "om__nails__tapping_____dglass-wood__inst1",
+      "audio_id": "O0UEHzbbhLk_3067_3156",
+      "task": "object_material_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "sm__tool__scratching__mic_____inst1",
+      "audio_id": "Z6DFyIu5sB8_11434_11622",
+      "task": "subject_material_mc",
+      "reference_correct_votes": 2,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_om__26f6hb8ZZmc_12021_12230",
+      "audio_id": "26f6hb8ZZmc_12021_12230",
+      "task": "object_material_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_om__AGBB5HAnAb0_8910_9323",
+      "audio_id": "AGBB5HAnAb0_8910_9323",
+      "task": "object_material_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_om__NkKyY3eqll0_8547_8644p597732426304",
+      "audio_id": "NkKyY3eqll0_8547_8644p597732426304",
+      "task": "object_material_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_om___Ue_uJMIjv8_5112_5348",
+      "audio_id": "_Ue_uJMIjv8_5112_5348",
+      "task": "object_material_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_om__dBjX6b5eQG4_6797_6947p016780045352",
+      "audio_id": "dBjX6b5eQG4_6797_6947p016780045352",
+      "task": "object_material_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_om__kOaWaYnqUsU_3709p4813605442178_3953p6012698412696",
+      "audio_id": "kOaWaYnqUsU_3709p4813605442178_3953p6012698412696",
+      "task": "object_material_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_om__mCM9lhHaYis_2446p9289795918367_2743",
+      "audio_id": "mCM9lhHaYis_2446p9289795918367_2743",
+      "task": "object_material_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_om__n_LZTPIcZiE_448_566",
+      "audio_id": "n_LZTPIcZiE_448_566",
+      "task": "object_material_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_om__nzNLs3h0D20_10664_11290",
+      "audio_id": "nzNLs3h0D20_10664_11290",
+      "task": "object_material_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_sm__dBjX6b5eQG4_5400p158276643991_5626p305941043084",
+      "audio_id": "dBjX6b5eQG4_5400p158276643991_5626p305941043084",
+      "task": "subject_material_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_sm__fxEpoqM6WiQ_13223_13285",
+      "audio_id": "fxEpoqM6WiQ_13223_13285",
+      "task": "subject_material_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_vb__-3hxDkxuyD4_5576_5580",
+      "audio_id": "-3hxDkxuyD4_5576_5580",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_vb__1UjpdWk1q6A_3533_3863",
+      "audio_id": "1UjpdWk1q6A_3533_3863",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_vb__26f6hb8ZZmc_12021_12230",
+      "audio_id": "26f6hb8ZZmc_12021_12230",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_vb__519DvEqtCJ0_0_215",
+      "audio_id": "519DvEqtCJ0_0_215",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_vb__519DvEqtCJ0_215_331p06920634920635",
+      "audio_id": "519DvEqtCJ0_215_331p06920634920635",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_vb__CSE-lUUD5fo_204_240p69179138321994",
+      "audio_id": "CSE-lUUD5fo_204_240p69179138321994",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_vb__CSE-lUUD5fo_242p69179138321994_428",
+      "audio_id": "CSE-lUUD5fo_242p69179138321994_428",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_vb__DNIfcQ3Gky8_6314_6566",
+      "audio_id": "DNIfcQ3Gky8_6314_6566",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 1
+    },
+    {
+      "question_id": "t2_vb__EvV73G3XPls_10431_10656",
+      "audio_id": "EvV73G3XPls_10431_10656",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_vb__GraCIkTvsok_1145_1177",
+      "audio_id": "GraCIkTvsok_1145_1177",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_vb__O0UEHzbbhLk_3067_3156",
+      "audio_id": "O0UEHzbbhLk_3067_3156",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_vb__O8yd1PrzuEQ_6017p221859410431_6108",
+      "audio_id": "O8yd1PrzuEQ_6017p221859410431_6108",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_vb__UQIEsGQjVjU_366_603",
+      "audio_id": "UQIEsGQjVjU_366_603",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_vb__X_en4yAd_EQ_1591_2026",
+      "audio_id": "X_en4yAd_EQ_1591_2026",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_vb__fxEpoqM6WiQ_8258_8425p98902494331",
+      "audio_id": "fxEpoqM6WiQ_8258_8425p98902494331",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 1
+    },
+    {
+      "question_id": "t2_vb__ge2Gbha2_yA_5006p548390022675_5073p941224489796",
+      "audio_id": "ge2Gbha2_yA_5006p548390022675_5073p941224489796",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_vb__mHi01HKazIs_864_869",
+      "audio_id": "mHi01HKazIs_864_869",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_vb__mHi01HKazIs_869_880",
+      "audio_id": "mHi01HKazIs_869_880",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_vb__mNggZ83n_LQ_1320_1595",
+      "audio_id": "mNggZ83n_LQ_1320_1595",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_vb__n-kxooJROJA_385_705",
+      "audio_id": "n-kxooJROJA_385_705",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_vb__nlNHkxiH8uk_919_1803",
+      "audio_id": "nlNHkxiH8uk_919_1803",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_vb__r7cW6PH2yKk_668_844",
+      "audio_id": "r7cW6PH2yKk_668_844",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_vb__sHLvsboM8I4_199_204",
+      "audio_id": "sHLvsboM8I4_199_204",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_vb__sIZaSUKwlwk_139_397",
+      "audio_id": "sIZaSUKwlwk_139_397",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "t2_vb__tAxvsG9s-TM_1622p0969614512471_1718p4450793650794",
+      "audio_id": "tAxvsG9s-TM_1622p0969614512471_1718p4450793650794",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 1
+    },
+    {
+      "question_id": "vb__combs__crusts__inst0",
+      "audio_id": "QYo1TMdO5Mk_825_916p6069841269841",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "vb__combs__crusts__inst1",
+      "audio_id": "vVgzpsYYpQ4_3885_3895",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "vb__combs__crusts__inst2",
+      "audio_id": "vVgzpsYYpQ4_3905_3915",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "vb__combs__crusts__inst3",
+      "audio_id": "vVgzpsYYpQ4_3925_3935",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "vb__combs__crusts__inst4",
+      "audio_id": "vVgzpsYYpQ4_3945_3955",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "vb__combs__crusts__inst5",
+      "audio_id": "vVgzpsYYpQ4_3965_3975",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "vb__fingernails__wood__inst1",
+      "audio_id": "ulrn5Wl5cbs_2999_3128p0491609977325",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "vb__fingers_____inst1",
+      "audio_id": "mCM9lhHaYis_1705p5738775510204_2064p017233560091",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "vb__fingers_____inst2",
+      "audio_id": "mCM9lhHaYis_2446p9289795918367_2743",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "vb__fingers_____inst3",
+      "audio_id": "nzNLs3h0D20_10664_11290",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "vb__fingers_____inst5",
+      "audio_id": "zIS3CI2mdiU_1073_1122",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "vb__hands__foam__inst1",
+      "audio_id": "_-NQyWZ6eq0_1173_1309p651895833",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "vb__hands__wrap__inst1",
+      "audio_id": "xlhhD8Bk0mg_3518_3840",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "vb__nails__wood__inst1",
+      "audio_id": "BI4VLglIZ1I_756_1114",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "vb__spoon__mic__inst1",
+      "audio_id": "fxEpoqM6WiQ_13223_13285",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 0
+    },
+    {
+      "question_id": "vb__tool__mic__inst3",
+      "audio_id": "OJUXupVz1Ms_7294_7315p71492063492",
+      "task": "verb_mc",
+      "reference_correct_votes": 3,
+      "silence_correct_votes": 1
+    }
+  ]
+}

metadata/train.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:50d706ff324050e9bd23b5fc3bfc58387e9f3d93b36de58899b2430584aa28b5
+size 16928716

release_manifest.json

@@ -0,0 +1,34 @@
+{
+  "dataset": "DeepASMR-NSpeech",
+  "prepared_at": "2026-09-20T04:09:01Z",
+  "dataset_card": "README.md",
+  "creator_map": {
+    "n_creators": 16
+  },
+  "train_label": {
+    "file": "train/train_label.json",
+    "n_rows": 72666,
+    "missing_audio": 0
+  },
+  "test_label": {
+    "file": "test/test_label.json",
+    "n_rows": 1163,
+    "missing_audio": 0
+  },
+  "svo_aqa": {
+    "file": "SVO-AQA/test_subset/bank.json",
+    "n_questions": 431,
+    "n_clips": 267,
+    "n_by_task": {
+      "object_material_mc": 105,
+      "subject_material_mc": 59,
+      "verb_mc": 267
+    },
+    "audio_dependent_subset": {
+      "file": "SVO-AQA/test_subset/audio_dependent_subset.v1.json",
+      "n_questions": 55,
+      "n_clips": 50
+    }
+  },
+  "vocab_reference": "vocab/closed_vocab_reference.md"
+}

train/train_label.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a01bad64fc966aac85d8b14e6525f8125b7c04b8797dc69f1dbde957e8f9f11f
+size 18120244

vocab/closed_vocab_reference.md

@@ -0,0 +1,134 @@
+# Closed vocabularies (reference)
+
+This file summarizes the **taxonomy** used by DeepASMR-NSpeech and SVO-AQA.
+Full CSV tables, alias maps, LLM prompts, and evaluation scripts are in the
+**GitHub code repository** (not duplicated here).
+
+---
+
+## Verb taxonomy · 37 verbs → 18 superclasses
+
+### Single-verb superclasses (1 verb each)
+
+- **bubbling** → `bubbling` — Continuous gentle gas bubbles in liquid or wet foam—not solid scraping or crushing.
+- **blowing** → `blowing` — Directed airflow or breath on a surface—no sustained solid–solid slide.
+- **sizzling** → `sizzling` — Heated oily food micro-pops from cooking—not dry flame crackle alone.
+- **crackling** → `crackling` — Flame, combustion, or dry thermal crackle—not candy pop or package snap.
+- **popping** → `popping` — Sharp discrete micro-pops (candy, bubble-wrap, gel-pod)—not continuous crackle or crush.
+- **rotating** → `rotating` — Object spins in place with rotation/mechanical whirr—not flip of pages/cards.
+- **flipping** → `flipping` — Repeated flip/turn cyclic contacts (cards, pages)—not continuous rotation hum.
+- **ticking** → `ticking` — Regular mechanical tick from clock/timer escapement—not keyboard typing.
+- **tinkling** → `tinkling` — Metal/glass/crystal parts of the prop itself lightly colliding or ringing—object-intrinsic crisp shimmer (like crackling is flame-intrinsic); not tool-on-surface scrape, pour/spill, or shaken-container rattle.
+
+### Multi-verb superclasses (within-class disambiguation applies)
+
+#### `friction_contact`
+- **Verbs:** `rubbing`, `scratching`, `brushing`, `rolling`
+- **Class cue:** Sustained solid–solid slide, bristle stroke, scratch, or roll—not peel/cut, crush, stir-in-cup, wet smear, or modeling-pack peel/scrape.
+- **Disambiguation** (rubbing vs scratching vs brushing vs rolling):
+  - rubbing: Two flat surfaces/pads slide with sustained friction (hands, paper, sponge, fabric).
+  - scratching: Sharp tip/nail high-frequency friction on a flat surface (earpick tip, nail on mesh).
+  - brushing: Many bristles/fibers stroke the surface (spoolie, hair brush, makeup brush).
+  - rolling: Sphere, marble, bead, or roller slides or rotates on a surface while contacting it.
+
+#### `impulsive_contact`
+- **Verbs:** `tapping`, `typing`, `dropping`
+- **Class cue:** Short discrete impacts: tap, key press, drop—NOT sustained slide or crush.
+- **Disambiguation** (tapping vs typing vs dropping):
+  - tapping: Single or few light discrete impacts.
+  - typing: Repeated key/button impacts in sequence.
+  - dropping: Object falls and hits a surface.
+
+#### `cutting_penetration`
+- **Verbs:** `scraping`, `cutting`, `tearing`, `piercing`
+- **Class cue:** Edge/tool removes material; blade cuts, tears, pierces, or scrapes layer off—including modeling pack / face mask peel or scrape from ear or mic—NOT whole-object crush or stir-in-container.
+- **Disambiguation** (scraping vs tearing vs cutting vs piercing):
+  - scraping: Hard edge/tool shaves or removes material from surface (dried modeling pack + metal tool).
+  - tearing: Whole layer/sheet ripped open by pulling force.
+  - cutting: Blade slices through material.
+  - piercing: Sharp pointed tip penetrates through material—distinct penetration sound (needle through fabric, tip through soft substance), not surface scrape or blade slice.
+
+#### `compression_deformation`
+- **Verbs:** `pressing`, `squeezing`, `crushing`, `smashing`, `stretching`
+- **Class cue:** Pressure deforms, stretches, or breaks a solid—crinkle, crush, smash—NOT blade peel, wet wipe, or stir inside cup.
+- **Disambiguation** (pressing vs squeezing vs crushing vs smashing vs stretching):
+  - pressing: Pressure without clear break (paper crinkling, light compress).
+  - squeezing: Enclosing soft deformable object.
+  - crushing: Hard material or crust collapses under strong force (floral foam, dried flowers, popcorn).
+  - smashing: Wet or sticky object being crushed (slime, jelly, floam, orbeez, water beads).
+  - stretching: Viscous slime/gel/cotton pulled apart under tension without full layer tear.
+
+#### `vocal_mouth`
+- **Verbs:** `whispering`, `crunching`, `licking`
+- **Class cue:** Whisper, chew/crunch, or lick—oral cavity only; no external tool on mic.
+- **Disambiguation** (whispering vs crunching vs licking):
+  - whispering: Breathy low inaudible speech.
+  - crunching: Mouth/teeth on food or brittle micro-breakage in oral/eating context.
+  - licking: Tongue/lip wet contact.
+
+#### `fluid_transfer`
+- **Verbs:** `pouring`, `spraying`
+- **Class cue:** Liquid or particle stream between containers or from nozzle—NOT smear on fixed surface.
+- **Disambiguation** (pouring vs spraying):
+  - pouring: Continuous liquid or granular stream transferred between containers or onto a surface.
+  - spraying: Atomized liquid from spray bottle, nozzle, or pressurized source—not a thick pour.
+
+#### `liquid_spread_wipe`
+- **Verbs:** `spreading`, `wiping`, `sticking`
+- **Class cue:** Wet/sticky medium smeared, wiped, or adhered on a surface—NOT dry modeling-pack peel/scrape, NOT in-cup stirring.
+- **Disambiguation** (spreading vs wiping vs sticking):
+  - spreading: Paste, cream, gel, or coating smeared/distributed across a surface (sticky wet drag).
+  - wiping: Broad pad, cloth, or hand removes or redistributes wet medium on a surface.
+  - sticking: Adhesive tape/sticker/paint tack attaches or detaches with sticky pull-release clicks.
+
+#### `container_agitation`
+- **Verbs:** `stirring`, `shaking`
+- **Class cue:** Stir or shake contents inside a container—NOT external slide on mic fur alone.
+- **Disambiguation** (stirring vs shaking):
+  - stirring: Circular or repeated agitation of liquid/granules inside a container (spoon, stick).
+  - shaking: Oscillatory shake causing contents to rattle, slosh, or impact container walls.
+
+#### `writing_painting`
+- **Verbs:** `writing`, `painting`
+- **Class cue:** Pen/stylus line-making or brush moving paint on a surface—not ear cleaning friction.
+- **Disambiguation** (writing vs painting):
+  - writing: Pen, pencil, marker, or stylus line-making friction on paper or hard surface.
+  - painting: Brush, palette knife, or tool applies/moves paint across a surface.
+
+---
+
+## Material taxonomy · 12 superclasses
+
+Used in SVO-AQA material multiple-choice questions (`object_material_mc`,
+`subject_material_mc`). Clip-level material labels are **not** shipped in
+train/test JSON; MC options carry the material gold labels.
+
+Note: the release name is **`clay`** (modeling pack / wet sticky sound),
+not `ceramic`.
+
+- **plastic** — stiff polymer body; crisp hollow resonance
+- **wood** — woody hardstock; dry grain friction
+- **clay** — modeling pack; wet sticky sound
+- **foam_lather** — wet liquid lather or shaving foam only; soft squish—not solid bar soap
+- **foam_solid** — expanded polymer foam; bead squeak porous sponge
+- **glass** — bright brittle crystal body or ice or ceramic; sharp ring transient
+- **paper** — light fibrous sheet; crinkle and tear rustle
+- **metal** — metallic solid or foil; bright ring transient
+- **rubber** — elastic silicone or latex; damped squeak stick-slip
+- **wax** — wax or solid bar soap; warm low-mid scrape tap or carve
+- **leather** — animal hide skin; flex creak surface rub
+- **textile_fibrous** — woven pile fabric; soft cotton wool rustle
+
+---
+
+## GitHub (full artifacts)
+
+For reproducibility, obtain from the project repository:
+
+- `vocab/asmr_verbs_en.csv` — 37 verb short hints
+- `vocab/asmr_verb_classes_en.csv` — verb → superclass
+- `vocab/asmr_material_classes_compact_en.csv` — 12 material classes
+- `vocab/stage2_class_disambiguation.txt` — fine-grained rules
+- `vocab/verb_alias.csv`, `vocab/material_alias.csv` — alias maps
+- Annotation prompts under `annotation/`
+