StyleInstructCaps commited on
Commit
563ce83
·
verified ·
1 Parent(s): 4ea4332

Add files using upload-large-folder tool

Browse files
README.md ADDED
@@ -0,0 +1,223 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc-by-nc-4.0
3
+ language:
4
+ - en
5
+ task_categories:
6
+ - text-to-speech
7
+ configs:
8
+ - config_name: StyleInstructCaps
9
+ data_files:
10
+ - split: train
11
+ path: data/StyleInstructCaps/train-*.parquet
12
+ - split: eval
13
+ path: data/StyleInstructCaps/eval-*.parquet
14
+ - config_name: OODPrompts
15
+ data_files:
16
+ - split: test
17
+ path: data/OODPrompts/test-*.parquet
18
+ - config_name: MetaSet
19
+ data_files:
20
+ - split: test
21
+ path: data/MetaSet/test-*.parquet
22
+ - config_name: IDBalanced
23
+ data_files:
24
+ - split: test
25
+ path: data/IDBalanced/test-*.parquet
26
+ - config_name: UDVCTK
27
+ data_files:
28
+ - split: test
29
+ path: data/UDVCTK/test-*.parquet
30
+ - config_name: UnseenEval
31
+ data_files:
32
+ - split: test
33
+ path: data/UnseenEval/test-*.parquet
34
+ ---
35
+
36
+ # StyleInstructCaps
37
+
38
+ Style-captioned speech metadata for speaking-style captioning and instruction-following evaluation, packaged in a [CapSpeech](https://huggingface.co/datasets/OpenSound/CapSpeech)-style Hugging Face dataset layout.
39
+
40
+ **Repository placeholder:** `<HF_ORG>/<DATASET_NAME>`
41
+
42
+ ## Overview
43
+
44
+ This dataset release contains style captions, transcriptions, tags, and instruction prompts for speech captioning research. It includes:
45
+
46
+ - **StyleInstructCaps** — large-scale training and evaluation splits with rich style annotations
47
+ - **OODPrompts** — out-of-distribution prompt evaluation on VCTK (32 utterances, 36 instruction prompts per row)
48
+ - **MetaSet** — evaluation subset derived from the Eval-500 benchmark (434 utterances with full style and instruction fields)
49
+ - **IDBalanced** — in-distribution, accent/gender-balanced evaluation set
50
+ - **UDVCTK** — VCTK evaluation with 10 speaker-idiosyncratic style prompts
51
+ - **UnseenEval** — unseen GigaSpeech-style evaluation set with ground-truth captions
52
+
53
+ This is a **metadata-only** release. Audio is **not** redistributed. Each row includes a `relative_audio_path` that should be resolved against the corresponding upstream audio corpus listed below.
54
+
55
+ During release preparation, local absolute `audio_path` values are removed, mixed string/list fields are normalized to consistent types, and numeric fields (`duration`, `snr`, `utterance_pitch_mean`) are stored as `float64`.
56
+
57
+ ## License
58
+
59
+ All resources are released under the [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/) license.
60
+
61
+ ## Usage
62
+
63
+ ```python
64
+ from datasets import load_dataset
65
+
66
+ REPO = "<HF_ORG>/<DATASET_NAME>"
67
+
68
+ # Load a specific config and split
69
+ train = load_dataset(REPO, "StyleInstructCaps", split="train")
70
+ eval_split = load_dataset(REPO, "StyleInstructCaps", split="eval")
71
+
72
+ # Load evaluation configs
73
+ ood = load_dataset(REPO, "OODPrompts", split="test")
74
+ metaset = load_dataset(REPO, "MetaSet", split="test")
75
+ id_balanced = load_dataset(REPO, "IDBalanced", split="test")
76
+ ud_vctk = load_dataset(REPO, "UDVCTK", split="test")
77
+ unseen = load_dataset(REPO, "UnseenEval", split="test")
78
+
79
+ # Inspect one example
80
+ print(train[0])
81
+ ```
82
+
83
+ ## Dataset Structure
84
+
85
+ Each config is stored as parquet files under `data/<config>/`. Local absolute `audio_path` values from the original manifests are **not** included; use `relative_audio_path` to locate audio in upstream datasets.
86
+
87
+ ### StyleInstructCaps
88
+
89
+ | Split | Rows | Description |
90
+ | ----- | ---- | ----------- |
91
+ | `train` | 1,031,840 | Full-scale training set with style captions and tags (11 parquet shards) |
92
+ | `eval` | 14,756 | Held-out evaluation set with style captions, tags, and instruction fields |
93
+
94
+ **Shared columns (train and eval)**
95
+
96
+ | Column | Type | Description |
97
+ | ------ | ---- | ----------- |
98
+ | `source` | string | Upstream dataset identifier (e.g., `emilia`, `voxceleb`, `ears`, `expresso`) |
99
+ | `relative_audio_path` | string | Relative path to the audio file within the upstream corpus |
100
+ | `transcription` | string | Transcript of the utterance |
101
+ | `text_description` | list[string] | Natural-language style caption(s) |
102
+ | `speaker_idiosyncratic_style` | string | Speaker identity style caption |
103
+ | `situational_contextual_style` | string | Situational/contextual style caption |
104
+ | `expressive_emotional_style` | string | Expressive/emotional style caption |
105
+ | `linguistic_pragmatic_style` | string | Linguistic/pragmatic style caption |
106
+ | `perceptual_listener_centric_style` | string | Listener-centric style caption |
107
+ | `holistic_creative_synthesis` | string | Holistic creative synthesis caption |
108
+ | `intrinsic_tags` | list[string] | Intrinsic style tags |
109
+ | `situational_tags` | list[string] | Situational tags (empty list when absent) |
110
+ | `basic_tags` | list[string] | Basic attribute tags |
111
+ | `all_tags` | list[string] | Combined tag set |
112
+ | `speakerid` | string | Speaker identifier |
113
+ | `name` | string | Speaker name (when available) |
114
+ | `duration` | float | Utterance duration in seconds |
115
+ | `gender` | string | Gender label |
116
+ | `accent` | string | Accent label (empty string when absent) |
117
+ | `pitch` | string | Pitch descriptor |
118
+ | `speaking_rate` | string | Speaking-rate descriptor |
119
+ | `noise` | string | Recording environment descriptor |
120
+ | `utterance_pitch_mean` | float | Mean pitch (Hz) |
121
+ | `snr` | float | Signal-to-noise ratio |
122
+ | `phonemes` | string | Phoneme sequence |
123
+ | `task` | string | Task label |
124
+
125
+ **Eval-only columns**
126
+
127
+ | Column | Type | Description |
128
+ | ------ | ---- | ----------- |
129
+ | `speaker_idiosyncratic_style_instruction` | string | Instruction for speaker style caption |
130
+ | `situational_contextual_style_instruction` | string | Instruction for situational style |
131
+ | `expressive_emotional_style_instruction` | string | Instruction for expressive style |
132
+ | `linguistic_pragmatic_style_instruction` | string | Instruction for linguistic style |
133
+ | `perceptual_listener_centric_style_instruction` | string | Instruction for listener-centric style |
134
+ | `holistic_creative_synthesis_instruction` | string | Instruction for holistic synthesis |
135
+
136
+ ### OODPrompts
137
+
138
+ | Split | Rows | Description |
139
+ | ----- | ---- | ----------- |
140
+ | `test` | 32 | VCTK OOD prompt evaluation (36 instruction prompts per utterance) |
141
+
142
+ **Core columns:** `source`, `relative_audio_path`, `speakerid`, `accent`, and `prompt1_speaker_idiosyncratic_style_instruction` through `prompt36_holistic_creative_synthesis_style_instruction`, covering speaker-idiosyncratic, situational, expressive, linguistic, perceptual, and holistic prompt families.
143
+
144
+ ### MetaSet
145
+
146
+ | Split | Rows | Description |
147
+ | ----- | ---- | ----------- |
148
+ | `test` | 434 | Eval-500 benchmark subset with full style annotations and instructions |
149
+
150
+ Schema matches **StyleInstructCaps/eval** (all shared style caption fields plus the six `*_instruction` columns).
151
+
152
+ ### IDBalanced
153
+
154
+ | Split | Rows | Description |
155
+ | ----- | ---- | ----------- |
156
+ | `test` | 400 | Accent/gender-balanced in-distribution evaluation set |
157
+
158
+ **Core columns:** `source`, `relative_audio_path`, `speakerid`, `accent`, `duration`, `snr`, and `prompt1_speaker_idiosyncratic_style_instruction` through `prompt10_speaker_idiosyncratic_style_instruction`.
159
+
160
+ ### UDVCTK
161
+
162
+ | Split | Rows | Description |
163
+ | ----- | ---- | ----------- |
164
+ | `test` | 320 | VCTK evaluation with 10 speaker-idiosyncratic prompts |
165
+
166
+ **Core columns:** `source`, `relative_audio_path`, `speakerid`, `accent`, and `prompt1_speaker_idiosyncratic_style_instruction` through `prompt10_speaker_idiosyncratic_style_instruction`.
167
+
168
+ ### UnseenEval
169
+
170
+ | Split | Rows | Description |
171
+ | ----- | ---- | ----------- |
172
+ | `test` | 600 | Unseen GigaSpeech-style evaluation set |
173
+
174
+ **Core columns**
175
+
176
+ | Column | Type | Description |
177
+ | ------ | ---- | ----------- |
178
+ | `source` | string | Upstream dataset identifier (`eval_giga`) |
179
+ | `relative_audio_path` | string | Relative path within the evaluation GigaSpeech subset |
180
+ | `ground_truth_caption` | string | Reference style caption |
181
+ | `holistic_creative_synthesis_instruction` | string | Instruction for holistic caption generation |
182
+
183
+ ## Audio Sources
184
+
185
+ Map each row's `source` and `relative_audio_path` to the corresponding upstream corpus:
186
+
187
+ | `source` value | Upstream audio corpus |
188
+ | -------------- | --------------------- |
189
+ | `emilia` | [Emilia-Dataset (EN)](https://huggingface.co/datasets/amphion/Emilia-Dataset) |
190
+ | `voxceleb` | [VoxCeleb / VoxCeleb2](https://www.robots.ox.ac.uk/~vgg/data/voxceleb/) |
191
+ | `VCTK` | [VCTK Corpus](https://datashare.ed.ac.uk/handle/10283/2950) |
192
+ | `ears` | [EARS Dataset](https://github.com/facebookresearch/ears_dataset) |
193
+ | `expresso` | [Expresso](https://github.com/facebookresearch/textlesslib/tree/main/examples/expresso/dataset) |
194
+ | `eval_giga` | [GigaSpeech](https://huggingface.co/datasets/speechcolab/gigaspeech) (evaluation subset) |
195
+
196
+ ## Preparing and Uploading (maintainers)
197
+
198
+ From the project root:
199
+
200
+ ```bash
201
+ python -m venv .venv
202
+ source .venv/bin/activate
203
+ pip install -r release/scripts/requirements.txt
204
+
205
+ python release/scripts/prepare_release.py
206
+
207
+ # Edit release/README.md: replace <HF_ORG>/<DATASET_NAME> and citation
208
+ huggingface-cli login
209
+ python release/scripts/upload_to_hf.py --org <HF_ORG> --name <DATASET_NAME>
210
+ ```
211
+
212
+ ## Citation
213
+
214
+ ```bibtex
215
+ @misc{styleinstructcaps2026,
216
+ title = {StyleInstructCaps: Style-Captioned Speech Annotations},
217
+ author = {<TODO: authors>},
218
+ year = {2026},
219
+ howpublished = {\url{https://huggingface.co/datasets/<HF_ORG>/<DATASET_NAME>}}
220
+ }
221
+ ```
222
+
223
+ Replace `<HF_ORG>`, `<DATASET_NAME>`, and the citation fields before publishing.
data/IDBalanced/test-00000-of-00001.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d6d8d24c6970b562eb177ba4905cd032030eaf0b2702707cc0c6c1656113fde3
3
+ size 33858
data/MetaSet/test-00000-of-00001.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5c67ce2666730b1460853011b8671efd44f75fee9e6d406a87ee30c91db011de
3
+ size 420418
data/OODPrompts/test-00000-of-00001.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:89d7ba31b77cb4256a1ef5abe9557dcc9660e24ae5d1a440ebc386f97170e888
3
+ size 64000
data/StyleInstructCaps/eval-00000-of-00001.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:17afd13d2949a63b6acacc44442ee097ab7a74c25558fe6a945c24bb045c42ab
3
+ size 13078194
data/StyleInstructCaps/train-00000-of-00011.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:223c3dea5a5ece9b3cd1c7aef6f05914f8d05eca01a836c3e3c28734c94e36ee
3
+ size 86122391
data/StyleInstructCaps/train-00001-of-00011.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b0a92b0d069c329d9d73719bca9d408010cc874ba4b0d0df085e3e058679d4d6
3
+ size 86173266
data/StyleInstructCaps/train-00002-of-00011.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:829152601149795b140217a6d0afab9c4347d12172ee6d80989e1409dea5d0ae
3
+ size 86277608
data/StyleInstructCaps/train-00003-of-00011.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:779b61855a633524b87608b40cd70511b6dbc46948b1046ede3d8a9fd6175d37
3
+ size 86168178
data/StyleInstructCaps/train-00004-of-00011.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d027615e332bb0a50b8d3fe25fe5669ac5d3a49bf3ea549558cb48391bc2bdd9
3
+ size 86306501
data/StyleInstructCaps/train-00005-of-00011.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:efdd068f79f3e1e5ce401598a7404480a1613e942b8c3d33400a4054557640c7
3
+ size 86234437
data/StyleInstructCaps/train-00006-of-00011.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:01beff6135c363cfd0d8ee0ed1f006c69248b367d587d89020fc82292c33391a
3
+ size 86234391
data/StyleInstructCaps/train-00007-of-00011.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d900689869b88e8c4845bbcc52533a71980c4763ed9df60b98a0ede98932c40e
3
+ size 86278309
data/StyleInstructCaps/train-00008-of-00011.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d24f307fbfe2198075d8476e85753995ebe26b7ac4f4e3fff4c6b7eb98219c37
3
+ size 86151181
data/StyleInstructCaps/train-00009-of-00011.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e3953342fadd91f93ae732be6e6e223b0fcfacb24fba5f2071cb1718869308b9
3
+ size 86057880
data/StyleInstructCaps/train-00010-of-00011.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:55d1fd91917a1d13fd960fe3c447b432b2e37e6c2c1ea933b9a40d8ec26bfbf8
3
+ size 27398529
data/UDVCTK/test-00000-of-00001.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:278a9678269c1b5e3b2f5f5156b7c2e2dd74b07dfbfcaadfaef03c0df117df14
3
+ size 24630
data/UnseenEval/test-00000-of-00001.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0df07446eed592dac46f6d4f6be2fd503ad205d35e87edb8a15f8ddba370ff70
3
+ size 61270
hf_release_for_spkcap_6e6bb298.plan.md ADDED
@@ -0,0 +1,148 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ name: HF release for spkcap
3
+ overview: Package the 7 source JSON manifests as a single Hugging Face dataset repo (CapSpeech-style) with 6 configs and parquet files under `/home/schand30/scratchbsisman/schand30/spkcap/release/`, drop local absolute `audio_path`s, and provide scripts to prepare and push the release without modifying any source files.
4
+ todos:
5
+ - id: scaffold
6
+ content: "Create release/ skeleton: scripts/, data/, requirements.txt, README.md placeholder"
7
+ status: pending
8
+ - id: prepare_script
9
+ content: Write release/scripts/prepare_release.py that converts the 7 JSONs into parquet shards under release/data/<config>/, dropping local audio_path
10
+ status: pending
11
+ - id: readme
12
+ content: Write release/README.md with the configs YAML front-matter and CapSpeech-style body (overview, usage, per-config schemas, audio source mapping)
13
+ status: pending
14
+ - id: upload_script
15
+ content: Write release/scripts/upload_to_hf.py wrapper around huggingface_hub.upload_large_folder, parameterized by org/name
16
+ status: pending
17
+ - id: user_run
18
+ content: User installs deps, runs prepare_release.py, fills in repo name, runs upload_to_hf.py
19
+ status: pending
20
+ isProject: false
21
+ ---
22
+
23
+ ## Goal
24
+
25
+ Release the seven manifests as a HF dataset repo modeled on [`OpenSound/CapSpeech`](https://huggingface.co/datasets/OpenSound/CapSpeech): one repo, multiple configs, parquet shards, metadata-only (no audio), with a placeholder repo name (`<HF_ORG>/<DATASET_NAME>`) the user fills in before pushing.
26
+
27
+ All work happens under `/home/schand30/scratchbsisman/schand30/spkcap/release/`. Source files in `data/data/` and `data/evaluation/data_input_dir/` are read-only.
28
+
29
+ ## Config / split layout
30
+
31
+ Each source JSON has an `{"annotation": [...]}` wrapper with heterogeneous per-file schemas, so we use one HF "config" per logical dataset (this is what enables HF's split selector UI like CapSpeech's).
32
+
33
+ - `StyleInstructCaps`
34
+ - `train` <- `data/data/train_fullscale_with_desc_v2.json` (1,031,840 rows, 3.3 GB)
35
+ - `eval` <- `data/data/evaluation_holdout_with_desc_llama.json` (14,756 rows)
36
+ - `OODPrompts`
37
+ - `test` <- `data/evaluation/data_input_dir/vctk_manifest_2utt_36prompts_unseen.json` (32 rows)
38
+ - `Eval500`
39
+ - `test` <- `data/evaluation/data_input_dir/tbd1_500_llama_global.json` (434 rows)
40
+ - `IDBalanced`
41
+ - `test` <- `data/evaluation/data_input_dir/balance_acccent_gender_eval_set2_with_fields.json` (400 rows)
42
+ - `UDVCTK`
43
+ - `test` <- `data/evaluation/data_input_dir/vctk_manifest_10prompts.json` (320 rows)
44
+ - `UnseenEval`
45
+ - `test` <- `data/evaluation/data_input_dir/giga_manifest.json` (600 rows)
46
+
47
+ Each row keeps all original fields except the local absolute `audio_path` (dropped). `relative_audio_path` is preserved as the portable identifier; the README documents which upstream dataset each `source` value maps to.
48
+
49
+ ## Output layout under `release/`
50
+
51
+ ```text
52
+ release/
53
+ ├── README.md # dataset card with YAML front-matter declaring configs
54
+ ├── scripts/
55
+ │ ├── requirements.txt # huggingface_hub, datasets, pyarrow, tqdm
56
+ │ ├── prepare_release.py # JSON -> parquet under release/data/<config>/
57
+ │ └── upload_to_hf.py # pushes release/ to <HF_ORG>/<DATASET_NAME>
58
+ └── data/
59
+ ├── StyleInstructCaps/
60
+ │ ├── train-00000-of-000NN.parquet # ~100k rows per shard => ~11 shards
61
+ │ └── eval-00000-of-00001.parquet
62
+ ├── OODPrompts/test-00000-of-00001.parquet
63
+ ├── Eval500/test-00000-of-00001.parquet
64
+ ├── IDBalanced/test-00000-of-00001.parquet
65
+ ├── UDVCTK/test-00000-of-00001.parquet
66
+ └── UnseenEval/test-00000-of-00001.parquet
67
+ ```
68
+
69
+ ## README.md front-matter (mirrors CapSpeech "data files via YAML" pattern)
70
+
71
+ ```yaml
72
+ ---
73
+ license: cc-by-nc-4.0
74
+ language:
75
+ - en
76
+ task_categories:
77
+ - text-to-speech
78
+ configs:
79
+ - config_name: StyleInstructCaps
80
+ data_files:
81
+ - split: train
82
+ path: data/StyleInstructCaps/train-*.parquet
83
+ - split: eval
84
+ path: data/StyleInstructCaps/eval-*.parquet
85
+ - config_name: OODPrompts
86
+ data_files:
87
+ - split: test
88
+ path: data/OODPrompts/test-*.parquet
89
+ - config_name: Eval500
90
+ data_files:
91
+ - split: test
92
+ path: data/Eval500/test-*.parquet
93
+ - config_name: IDBalanced
94
+ data_files:
95
+ - split: test
96
+ path: data/IDBalanced/test-*.parquet
97
+ - config_name: UDVCTK
98
+ data_files:
99
+ - split: test
100
+ path: data/UDVCTK/test-*.parquet
101
+ - config_name: UnseenEval
102
+ data_files:
103
+ - split: test
104
+ path: data/UnseenEval/test-*.parquet
105
+ ---
106
+ ```
107
+
108
+ The README body mirrors CapSpeech's structure: Overview, License (CC BY-NC 4.0 placeholder), Usage snippet (`load_dataset("<HF_ORG>/<DATASET_NAME>", "StyleInstructCaps", split="train")`), per-config Schema tables, Audio Sources table mapping each `source` value (`emilia`, `voxceleb`, `VCTK`, `ears`, `eval_giga`) to its upstream dataset URL, and a Citation block with `<TODO citation>`.
109
+
110
+ ## `prepare_release.py` behavior
111
+
112
+ Single, idempotent script that:
113
+
114
+ 1. Loads each source JSON via `json.load(...)["annotation"]`. Train file (3.3 GB) needs ~10-12 GB RAM at peak; the script logs a warning and recommends running on a node with >=16 GB free RAM. (If that's a concern we can swap to `ijson` streaming, but starts simpler.)
115
+ 2. Drops the `audio_path` key from every record (keeps `relative_audio_path`).
116
+ 3. Builds a `datasets.Dataset` per file (preserves nested lists/nullable fields like `situational_tags`).
117
+ 4. Writes parquet to the layout above. For `StyleInstructCaps/train`, shards via `Dataset.shard(num_shards=ceil(N/100_000), index=i).to_parquet(...)` so each shard is roughly 100k rows / a few hundred MB. All other splits write a single `*-00000-of-00001.parquet`.
118
+ 5. Prints final row counts to verify against the source counts (`1031840 / 14756 / 32 / 434 / 400 / 320 / 600`).
119
+
120
+ ## `upload_to_hf.py` behavior
121
+
122
+ Thin wrapper around `huggingface_hub.HfApi`:
123
+
124
+ - Reads `HF_ORG` and `DATASET_NAME` from CLI args (or env).
125
+ - `create_repo(repo_id=f"{HF_ORG}/{DATASET_NAME}", repo_type="dataset", exist_ok=True)`.
126
+ - `upload_large_folder(folder_path="release/", repo_id=..., repo_type="dataset", ignore_patterns=["scripts/__pycache__/*"])`.
127
+ - Prints the resulting URL.
128
+
129
+ ## Step-by-step the user runs
130
+
131
+ 1. Create venv and install deps
132
+ - `python -m venv .venv && source .venv/bin/activate`
133
+ - `pip install -r release/scripts/requirements.txt`
134
+ 2. Generate the release artifacts
135
+ - `python release/scripts/prepare_release.py`
136
+ - Verify `release/data/...` shape and row counts in stdout.
137
+ 3. Edit `release/README.md`: fill in `<HF_ORG>/<DATASET_NAME>`, citation, license confirmation, and any per-config notes.
138
+ 4. Authenticate with HF: `huggingface-cli login` (paste a write-scoped token).
139
+ 5. Push:
140
+ - `python release/scripts/upload_to_hf.py --org <HF_ORG> --name <DATASET_NAME>`
141
+ - Or equivalently: `huggingface-cli upload <HF_ORG>/<DATASET_NAME> release/ . --repo-type=dataset`.
142
+ 6. On the HF dataset page, confirm the config/split selector shows all 7 splits and that the data viewer renders rows.
143
+
144
+ ## Things explicitly NOT done
145
+
146
+ - No edits to any source JSON in `data/data/` or `data/evaluation/data_input_dir/`.
147
+ - No upload of audio files; users are pointed at upstream datasets for the audio.
148
+ - Repo name kept as a placeholder until the user provides it.
scripts/prepare_release.py ADDED
@@ -0,0 +1,295 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python3
2
+ """Convert source JSON manifests into Hugging Face release parquet files.
3
+
4
+ Reads source files under data/ (read-only), drops local absolute audio_path,
5
+ and writes parquet shards under release/data/<config>/.
6
+ """
7
+
8
+ from __future__ import annotations
9
+
10
+ import argparse
11
+ import json
12
+ import sys
13
+ from pathlib import Path
14
+ from typing import Any, Iterator
15
+
16
+ import ijson
17
+ from datasets import Dataset
18
+ from tqdm import tqdm
19
+
20
+ PROJECT_ROOT = Path(__file__).resolve().parents[2]
21
+ RELEASE_ROOT = PROJECT_ROOT / "release"
22
+ DATA_ROOT = RELEASE_ROOT / "data"
23
+
24
+ ROWS_PER_SHARD = 100_000
25
+
26
+ # Expected row counts for verification.
27
+ EXPECTED_COUNTS = {
28
+ ("StyleInstructCaps", "train"): 1_031_840,
29
+ ("StyleInstructCaps", "eval"): 14_756,
30
+ ("OODPrompts", "test"): 32,
31
+ ("MetaSet", "test"): 434,
32
+ ("IDBalanced", "test"): 400,
33
+ ("UDVCTK", "test"): 320,
34
+ ("UnseenEval", "test"): 600,
35
+ }
36
+
37
+ RELEASE_SPECS: list[dict[str, Any]] = [
38
+ {
39
+ "config": "StyleInstructCaps",
40
+ "split": "train",
41
+ "source": PROJECT_ROOT / "data/data/train_fullscale_with_desc_v2.json",
42
+ "stream": True,
43
+ },
44
+ {
45
+ "config": "StyleInstructCaps",
46
+ "split": "eval",
47
+ "source": PROJECT_ROOT / "data/data/evaluation_holdout_with_desc_llama.json",
48
+ "stream": False,
49
+ },
50
+ {
51
+ "config": "OODPrompts",
52
+ "split": "test",
53
+ "source": PROJECT_ROOT
54
+ / "data/evaluation/data_input_dir/vctk_manifest_2utt_36prompts_unseen.json",
55
+ "stream": False,
56
+ },
57
+ {
58
+ "config": "MetaSet",
59
+ "split": "test",
60
+ "source": PROJECT_ROOT
61
+ / "data/evaluation/data_input_dir/tbd1_500_llama_global.json",
62
+ "stream": False,
63
+ },
64
+ {
65
+ "config": "IDBalanced",
66
+ "split": "test",
67
+ "source": PROJECT_ROOT
68
+ / "data/evaluation/data_input_dir/balance_acccent_gender_eval_set2_with_fields.json",
69
+ "stream": False,
70
+ },
71
+ {
72
+ "config": "UDVCTK",
73
+ "split": "test",
74
+ "source": PROJECT_ROOT
75
+ / "data/evaluation/data_input_dir/vctk_manifest_10prompts.json",
76
+ "stream": False,
77
+ },
78
+ {
79
+ "config": "UnseenEval",
80
+ "split": "test",
81
+ "source": PROJECT_ROOT
82
+ / "data/evaluation/data_input_dir/giga_manifest.json",
83
+ "stream": False,
84
+ },
85
+ ]
86
+
87
+
88
+ STYLE_TEXT_FIELDS = {
89
+ "speaker_idiosyncratic_style",
90
+ "situational_contextual_style",
91
+ "expressive_emotional_style",
92
+ "linguistic_pragmatic_style",
93
+ "perceptual_listener_centric_style",
94
+ "holistic_creative_synthesis",
95
+ "captions",
96
+ }
97
+
98
+ LIST_FIELDS = {
99
+ "text_description",
100
+ "intrinsic_tags",
101
+ "situational_tags",
102
+ "basic_tags",
103
+ "all_tags",
104
+ }
105
+
106
+
107
+ def _to_text(value: Any) -> str | None:
108
+ if value is None:
109
+ return None
110
+ if isinstance(value, str):
111
+ return value
112
+ if isinstance(value, list):
113
+ parts = [str(item) for item in value if item is not None]
114
+ return " ".join(parts) if parts else None
115
+ return str(value)
116
+
117
+
118
+ def _to_string_list(value: Any) -> list[str]:
119
+ if value is None:
120
+ return []
121
+ if isinstance(value, str):
122
+ return [value]
123
+ if isinstance(value, list):
124
+ return [str(item) for item in value if item is not None]
125
+ return [str(value)]
126
+
127
+
128
+ FLOAT_FIELDS = {
129
+ "duration",
130
+ "snr",
131
+ "utterance_pitch_mean",
132
+ }
133
+
134
+
135
+ def _to_float(value: Any) -> float | None:
136
+ if value is None:
137
+ return None
138
+ return float(value)
139
+
140
+
141
+ def normalize_record(record: dict[str, Any]) -> dict[str, Any]:
142
+ """Normalize heterogeneous JSON values to parquet-friendly types."""
143
+ cleaned = dict(record)
144
+ cleaned.pop("audio_path", None)
145
+
146
+ for field in STYLE_TEXT_FIELDS:
147
+ if field in cleaned:
148
+ cleaned[field] = _to_text(cleaned[field])
149
+
150
+ for field in LIST_FIELDS:
151
+ if field in cleaned:
152
+ cleaned[field] = _to_string_list(cleaned[field])
153
+
154
+ if "accent" in cleaned and cleaned["accent"] is None:
155
+ cleaned["accent"] = ""
156
+
157
+ for field in FLOAT_FIELDS:
158
+ if field in cleaned:
159
+ cleaned[field] = _to_float(cleaned[field])
160
+
161
+ return cleaned
162
+
163
+
164
+ def sanitize_record(record: dict[str, Any]) -> dict[str, Any]:
165
+ """Drop local absolute audio_path and normalize field types."""
166
+ return normalize_record(record)
167
+
168
+
169
+ def iter_annotation_records(source_path: Path) -> Iterator[dict[str, Any]]:
170
+ """Stream records from a large JSON file with an annotation array."""
171
+ with source_path.open("rb") as handle:
172
+ for record in ijson.items(handle, "annotation.item"):
173
+ yield sanitize_record(record)
174
+
175
+
176
+ def load_annotation_records(source_path: Path) -> list[dict[str, Any]]:
177
+ with source_path.open("r", encoding="utf-8") as handle:
178
+ payload = json.load(handle)
179
+ records = payload["annotation"]
180
+ return [sanitize_record(record) for record in records]
181
+
182
+
183
+ def write_single_parquet(
184
+ records: list[dict[str, Any]],
185
+ output_dir: Path,
186
+ split: str,
187
+ ) -> int:
188
+ output_dir.mkdir(parents=True, exist_ok=True)
189
+ dataset = Dataset.from_list(records)
190
+ output_path = output_dir / f"{split}-00000-of-00001.parquet"
191
+ dataset.to_parquet(str(output_path))
192
+ return len(records)
193
+
194
+
195
+ def write_sharded_parquet(
196
+ record_iter: Iterator[dict[str, Any]],
197
+ output_dir: Path,
198
+ split: str,
199
+ rows_per_shard: int,
200
+ ) -> int:
201
+ """Write parquet shards from a streaming iterator."""
202
+ output_dir.mkdir(parents=True, exist_ok=True)
203
+
204
+ total_rows = 0
205
+ shard_rows: list[dict[str, Any]] = []
206
+ shard_paths: list[Path] = []
207
+
208
+ def flush_shard(rows: list[dict[str, Any]], index: int) -> None:
209
+ if not rows:
210
+ return
211
+ path = output_dir / f"{split}-shard-{index:05d}.parquet"
212
+ Dataset.from_list(rows).to_parquet(str(path))
213
+ shard_paths.append(path)
214
+
215
+ for record in tqdm(record_iter, desc=f"{output_dir.name}/{split}", unit="rows"):
216
+ shard_rows.append(record)
217
+ total_rows += 1
218
+ if len(shard_rows) >= rows_per_shard:
219
+ flush_shard(shard_rows, len(shard_paths))
220
+ shard_rows = []
221
+
222
+ if shard_rows:
223
+ flush_shard(shard_rows, len(shard_paths))
224
+
225
+ total_shards = len(shard_paths)
226
+ for index, shard_path in enumerate(shard_paths):
227
+ final_name = output_dir / f"{split}-{index:05d}-of-{total_shards:05d}.parquet"
228
+ shard_path.rename(final_name)
229
+
230
+ return total_rows
231
+
232
+
233
+ def process_spec(spec: dict[str, Any], rows_per_shard: int) -> tuple[str, str, int]:
234
+ config = spec["config"]
235
+ split = spec["split"]
236
+ source: Path = spec["source"]
237
+ output_dir = DATA_ROOT / config
238
+
239
+ if not source.exists():
240
+ raise FileNotFoundError(f"Missing source file: {source}")
241
+
242
+ print(f"\nProcessing {config}/{split} from {source}")
243
+
244
+ if spec.get("stream"):
245
+ print(
246
+ " Streaming large file with ijson (~3.3 GB JSON). "
247
+ "Recommend >=16 GB free RAM for parquet conversion."
248
+ )
249
+ row_count = write_sharded_parquet(
250
+ iter_annotation_records(source),
251
+ output_dir=output_dir,
252
+ split=split,
253
+ rows_per_shard=rows_per_shard,
254
+ )
255
+ else:
256
+ records = load_annotation_records(source)
257
+ row_count = write_single_parquet(records, output_dir=output_dir, split=split)
258
+
259
+ expected = EXPECTED_COUNTS.get((config, split))
260
+ if expected is not None and row_count != expected:
261
+ print(
262
+ f" WARNING: row count mismatch for {config}/{split}: "
263
+ f"got {row_count}, expected {expected}",
264
+ file=sys.stderr,
265
+ )
266
+ else:
267
+ print(f" Wrote {row_count} rows to {output_dir}")
268
+
269
+ return config, split, row_count
270
+
271
+
272
+ def main() -> None:
273
+ parser = argparse.ArgumentParser(description="Prepare Hugging Face release parquet files.")
274
+ parser.add_argument(
275
+ "--rows-per-shard",
276
+ type=int,
277
+ default=ROWS_PER_SHARD,
278
+ help="Rows per parquet shard for the large train split.",
279
+ )
280
+ args = parser.parse_args()
281
+
282
+ print(f"Project root: {PROJECT_ROOT}")
283
+ print(f"Release data dir: {DATA_ROOT}")
284
+
285
+ summary: list[tuple[str, str, int]] = []
286
+ for spec in RELEASE_SPECS:
287
+ summary.append(process_spec(spec, rows_per_shard=args.rows_per_shard))
288
+
289
+ print("\nSummary:")
290
+ for config, split, count in summary:
291
+ print(f" {config}/{split}: {count:,} rows")
292
+
293
+
294
+ if __name__ == "__main__":
295
+ main()
scripts/requirements.txt ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ huggingface_hub>=0.26.0
2
+ datasets>=3.0.0
3
+ fsspec[http]>=2023.1.0,<=2024.9.0
4
+ pyarrow>=17.0.0
5
+ tqdm>=4.66.0
6
+ ijson>=3.3.0
scripts/upload_to_hf.py ADDED
@@ -0,0 +1,67 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python3
2
+ """Upload the prepared release folder to Hugging Face Hub."""
3
+
4
+ from __future__ import annotations
5
+
6
+ import argparse
7
+ import os
8
+ from pathlib import Path
9
+
10
+ from huggingface_hub import HfApi
11
+
12
+ PROJECT_ROOT = Path(__file__).resolve().parents[2]
13
+ RELEASE_ROOT = PROJECT_ROOT / "release"
14
+
15
+
16
+ def parse_args() -> argparse.Namespace:
17
+ parser = argparse.ArgumentParser(description="Upload release/ to Hugging Face Hub.")
18
+ parser.add_argument(
19
+ "--org",
20
+ default=os.environ.get("HF_ORG"),
21
+ help="Hugging Face organization or username (or set HF_ORG).",
22
+ )
23
+ parser.add_argument(
24
+ "--name",
25
+ default=os.environ.get("DATASET_NAME"),
26
+ help="Dataset repository name (or set DATASET_NAME).",
27
+ )
28
+ parser.add_argument(
29
+ "--release-dir",
30
+ type=Path,
31
+ default=RELEASE_ROOT,
32
+ help="Path to the release directory to upload.",
33
+ )
34
+ return parser.parse_args()
35
+
36
+
37
+ def main() -> None:
38
+ args = parse_args()
39
+ if not args.org or not args.name:
40
+ raise SystemExit(
41
+ "Missing repo identifiers. Provide --org and --name, "
42
+ "or set HF_ORG and DATASET_NAME environment variables."
43
+ )
44
+
45
+ release_dir = args.release_dir.resolve()
46
+ if not release_dir.exists():
47
+ raise SystemExit(f"Release directory not found: {release_dir}")
48
+
49
+ repo_id = f"{args.org}/{args.name}"
50
+ api = HfApi()
51
+
52
+ print(f"Creating dataset repo (if needed): {repo_id}")
53
+ api.create_repo(repo_id=repo_id, repo_type="dataset", exist_ok=True)
54
+
55
+ print(f"Uploading {release_dir} -> {repo_id}")
56
+ api.upload_large_folder(
57
+ folder_path=str(release_dir),
58
+ repo_id=repo_id,
59
+ repo_type="dataset",
60
+ ignore_patterns=["scripts/__pycache__/*", "**/.venv/**", "**/__pycache__/**"],
61
+ )
62
+
63
+ print(f"Upload complete: https://huggingface.co/datasets/{repo_id}")
64
+
65
+
66
+ if __name__ == "__main__":
67
+ main()