子集 (config)	split	檔數	筆數	音檔總長	大小	每筆訊息數	一句話說明
`dialogue`	train	96	48,000	778.3 h	20.4 GB	2–25（多為 7–11）	多輪情境對話，口語填充詞＋使用者插話（overlap）
	validation	10	4,791	78.2 h	2.1 GB
`control`	train	16	8,000	62.7 h	1.6 GB	3–9（多為 7）	使用者每輪下語氣/語速指令，助理照做
	validation	2	1,000	8.0 h	0.2 GB
`para_breezy`	train	19	6,507	116.5 h	17.9 GB	5 或 7	情境對話中途要求切換情緒/語速/耳語
	validation	5	1,693	31.1 h	4.8 GB
`if_48k`	train	11	44,000	69.7 h	11.0 GB	2	單輪指令遵循（朗讀、數數、拼字…），部分帶情緒/語速
	validation	1	4,000	6.3 h	1.0 GB
合計		160	117,991	1,150.8 h	58.9 GB

音檔總長以每筆最後一輪 timestamp_range 的結束時間加總。

原始來源

子集	原始路徑
`dialogue`	`TASTE2-SFT-Data/stage3/stage3_sft_*.parquet`（validation = 00097–00106）
`control`	`TASTE2-SFT-Data/stage3/control_stage3_sft_*.parquet`（validation = 00017、00018）
`para_breezy`	`output_parquet_breezy/part_*.parquet`（validation = 0000、0005、0010、0015、0020）
`if_48k`	`TASTE-IF-SFT-48K/shuffled_train_part_*.parquet`、`shuffled_dev.parquet`（validation）

檔名保留原始名稱，只按 split 分資料夾。

1. 文字在哪裡

沒有獨立的文字檔或 transcript 檔，所有文字都和音檔一起存在 parquet 裡。4 個子集的 schema 不同，尤其 if_48k 的 system prompt 位置不一樣：

想要的文字	`dialogue` / `control`	`para_breezy`	`if_48k`
System prompt	`message[0].text`（`role == "system"`）	`message[0].text`（`role == "system"`）	獨立欄位 `system_prompt`（`message` 裡沒有 system）
使用者說的話	`message[i].text`，`role == "user"`	同左	同左
助理回覆	`message[i].text`，`role == "assistant"`	同左	同左
語音風格標籤	無	`message[i].para_tags`（struct）	`message[i].para_tags`（JSON 字串）
類別	`meta.scenario`（全為 `dialogue`）、`idx` 前綴	`meta.para_dimension`、`meta.para_value`、`meta.topic`	`meta.instruction_type`

每一輪的 text 就是同一輪 audio 的內容。

2. Parquet 資料結構

2.1 `dialogue`、`control`

idx:  string
meta: struct<
        scenario: string                  # 固定為 "dialogue"
      >
message: list<struct<
        audio:           binary,          # 該輪完整 MP3 檔位元組；system 為 null
        role:            string,          # "system" | "user" | "assistant"
        text:            string,
        timestamp_range: list<int64>      # [開始ms, 結束ms]；system 為 null
      >>

2.2 `para_breezy`

idx:  string
meta: struct<
        conversation_id: string,          # 與 idx 相同
        mode:            string,          # 固定為 "para_breezy"
        topic:           string,          # 41 個主題（小寫）
        scenario:        string,          # 空字串
        para_dimension:  string,          # "emotion" | "speed" | "whisper"
        para_value:      string           # 目標風格，見 §3.3
      >
message: list<struct<
        role:            string,
        text:            string,
        audio:           binary,          # 該輪完整 WAV 檔位元組；system 為 null
        timestamp_range: list<int64>,
        speaker:         string,          # user 輪為 "user"，assistant 輪為 "agent"，system 為 null
        para_tags:       struct<          # 每個鍵沒有值時為 null
          gender: string, age: string, pitch: string,
          speed: string, volume: string, emotion: string
        >
      >>

2.3 `if_48k`

idx:           string
system_prompt: string                     # system prompt 在這裡，不在 message 裡
meta: struct<
        instruction_type: string          # 26 種指令類型，見 §3.4
      >
message: list<struct<                     # 固定 2 則：user, assistant
        role:            string,
        text:            string,
        audio:           binary,          # 該輪完整 WAV 檔位元組
        timestamp_range: list<int64>,
        para_tags:       string           # JSON 字串，例如 "{\"speed\": \"slow\"}"；沒有時為 null
      >>

欄位共通說明

欄位	說明
`message[].audio`	一個完整音檔，可直接寫到磁碟或用 `soundfile.read(io.BytesIO(b))` 讀取
`message[].timestamp_range`	毫秒，表示該輪在整段對話時間軸上的位置。把每輪音檔放在 `start` 位置即可還原整段對話。`dialogue` 中下一輪的開始可能早於上一輪的結束（插話重疊）

音檔格式（各檔第一筆抽查）

子集	格式
`dialogue`、`control`	MP3（MPEG-2 Layer III），24 kHz，單聲道，64 kbps
`para_breezy`、`if_48k`	WAV PCM int16，24 kHz，單聲道

混用時請注意 dialogue / control 是 MP3，不是 WAV。

3. 各子集說明與文字範例

3.1 `dialogue`：多輪情境對話（含填充詞與插話）

合成的英文多輪對話，涵蓋客服、技術支援、購物、旅遊等情境。system prompt 有 24,692 種，大部分是角色設定，例如 "You are a technical support specialist, guide the user through..."；另有 1,970 筆是 "You are a helpful assistant."，600 筆是 "You are a helpful assistant in a dialogue scenario."。

idx 格式為 <前綴>.<資料批次>.scenario<n>_<k>，前綴代表生成設定：

`idx` 前綴	筆數	使用者有口語填充詞 (um/uh/like…)	有插話重疊	assistant 被打斷（句尾為「—」）
`syn_data`	28,975	100%	85%	30%
`syn_data_n_filler`	15,890	35%	98%	21%
`syn_data_n_filler_n_overlap`	7,926	23%	0%	0%

（n_filler = 少/無填充詞，n_overlap = 無重疊。百分比是用關鍵字與時間戳估算的。）

插話：使用者那輪的 timestamp_range[0] 早於上一輪 assistant 的結束時間，被打斷的 assistant 文字以「—」結尾。
17,040 筆以 user 輪結尾。

idx: syn_data.data1.scenario4_10

[SYSTEM]    As a supply chain specialist, discuss the importance of accurate packaging labels, proper packaging materials, and clear communication with shipping carriers to minimize the risk of packages going missing during transit...
[USER]      [0 → 10789]       I'm, um, hoping you can help me find my missing package, I've been waiting for weeks and I'm—well, kinda starting to get worried. Can, uh, can you please assist me?
[ASSISTANT] [10789 → 17300]   I'd be happy to help you track down your package. Can you please provide me with the tracking number?
[USER]      [17300 → 26155]   Yeah, yeah. It's #HY456789US, uh, I've been checking the website, but it's—it's not giving me any updates.
[ASSISTANT] [26155 → 31791]   Let me see what I can find out... it looks like the package was delayed due to customs clearance.
...

插話範例（下一輪 user 的開始時間 39151 早於 assistant 的結束時間 40691）：

idx: syn_data.data4-2.scenario25_16

[ASSISTANT] [31190 → 40691]   Sure, one red flag might be programs that make extreme promises like "lose 20 pounds in one week." Since sustainable progress takes time, extreme claims—
[USER]      [39151 → 44713]   Aren't those usually, hmm, crash diets or really harmful approaches though?

3.2 `control`：逐輪語氣/語速指令

使用者每一輪都下一個語氣或語速指令，助理用簡短的話回應，並維持該風格直到下一個指令。
system prompt 共 5 種，都在說明「依使用者指令調整語氣與語速」，例如：
- "You are a performer agent. Treat User input as directions for tone and pace. Adjust to the latest instruction and maintain that style until changed, keeping responses smooth and natural."
- "You are a control-following agent. Always obey tone and pace commands, applying the latest instruction until it changes. Default to neutral if no instruction is given."
**沒有 para_tags**，指令只存在於使用者文字中。
idx 格式為 control.control_txt_<n>.scenarios_<k>。

idx: control.control_txt_1.scenarios_1

[USER]      Can you speak in a calm tone?
[ASSISTANT] Of course, I'm speaking calmly now.
[USER]      Now, switch to a surprised tone.
[ASSISTANT] Oh, wow, I'm surprised by your request.
...

idx: control.control_txt_30.scenarios_101

[USER]      Can you speak fast?
[ASSISTANT] Ok, I am speaking fast.
[USER]      Now, speak in a lonely tone.
[ASSISTANT] I feel very lonely.
[USER]      Can you also sound surprised?
...

3.3 `para_breezy`：情境對話中途切換風格

41 個主題 × 200 筆：art, books, cars, celebrities, coding, cooking, education, events, fashion, finance, fitness, food, gaming, gardening, health, history, hobbies, holidays, home, languages, makeup, movies, music, nature, news, pets, philosophy, photography, podcasts, politics, relationships, science, shopping, social_media, spirituality, sports, technology, traditions, travel, weather, work。

每個情境做成 10 種目標風格，各 820 筆：

`meta.para_dimension`	`meta.para_value`	被標記的 assistant 輪的 `para_tags`
`emotion`	`angry` `disgusted` `fear` `happy` `neutral` `sad` `surprise`	`emotion = <值>`
`speed`	`fast` `slow`	`speed = <值>`
`whisper`	`whisper`	`volume = "whisper"`（注意是放在 `volume` 鍵）

System prompt 全部相同："You are a helpful conversational assistant. Respond naturally, adapting your tone to match the user's emotional state."
第 2 則 user（message[3]）會用文字提出風格要求，只有緊接著的那一則 assistant（message[4]）帶 para_tags，其餘輪的 para_tags 欄位全為 null。
idx 格式為 <topic>_scenario<n>_<para_dimension>_<para_value>。

idx: books_scenario5_speed_fast    meta.para_dimension: speed    meta.para_value: fast

[SYSTEM]    You are a helpful conversational assistant. Respond naturally, adapting your tone to match the user's emotional state.
[USER]      I'm trying to self-publish my first book, but I have no idea where to start, can you help me with that?
[ASSISTANT] Absolutely, I'd love to help! Start by writing and finalizing your manuscript. Then, focus on formatting it for both e-book and print...
[USER]      Could you speed up a little? I'm interested in the formatting part, what are some common tools that authors use for that?
[ASSISTANT] Sure! Common tools for formatting include Scrivener, Vellum (for Mac), and Reedsy...      para_tags.speed = "fast"
[USER]      That's really helpful, yeah, I've heard of Scrivener, can you tell me more about Vellum and Reedsy...?
[ASSISTANT] Vellum is great for Mac users, offering a user-friendly interface and beautiful formatting options...

3.4 `if_48k`：單輪指令遵循

每筆固定 2 則訊息：user → assistant。
System prompt 存在 system_prompt 欄位，全部相同："You are a helpful spoken assistant. Listen to the user's spoken instruction and respond with only the requested spoken content, following the instruction exactly."
meta.instruction_type 共 26 種（train + validation）：

read_aloud 37,308、counting 1,240、repetition 1,133、spelling 982、sequence 846、negative_constraint 792、number_reading 774、format_constraint 752、multi_step 704、listing 514、conditional 406、exact_count 347、simple_arithmetic 294、time_date_reading 207、word_extraction 204、required_word 200、replacement 200、short_description 164、completion 162、comparison 154、reverse_sequence 142、filtering 141、short_generation 111、ordering 110、selection 72、transformation 41
只有 assistant 輪會帶 para_tags：emotion（happy / sad / angry / surprised / fearful / disgusted / whisper）或 speed（fast / slow）。

{
  "idx": "read_aloud_016661",
  "system_prompt": "You are a helpful spoken assistant. Listen to the user's spoken instruction and respond with only the requested spoken content, following the instruction exactly.",
  "meta": {"instruction_type": "read_aloud"},
  "message": [
    {"role": "user", "text": "Say each number from one to five slowly.", "audio": "<wav bytes>", "timestamp_range": [0, 3030], "para_tags": null},
    {"role": "assistant", "text": "One, two, three, four, five.", "audio": "<wav bytes>", "timestamp_range": [3030, 7914], "para_tags": "{\"speed\": \"slow\"}"}
  ]
}

其他範例：

[USER]      Please spell the word table backwards.                 (spelling)
[ASSISTANT] E, L, B, A, T.

[USER]      Choose the correct tool: hammer, orange, or pillow.    (selection)
[ASSISTANT] Hammer.

4. 已知問題

para_breezy 有 513 筆的風格要求文字與標籤不符。 有 57 個情境的 9 種非 angry 版本（disgusted / fear / happy / neutral / sad / surprise / fast / slow / whisper 各 57 筆），使用者的要求文字仍是 "Can you speak in an angry tone…"，但 para_value 與 assistant 的 para_tags 是其他風格。例：books_scenario13_emotion_happy、books_scenario4_whisper_whisper。
dialogue 有 12 筆沒有 user 輪（只有 system + assistant），例如 syn_data.data2.scenario33_4。
dialogue 有 17,040 筆、control 有 19 筆以 user 輪結尾。 訓練時若需要以 assistant 結尾，請自行截掉最後一輪。
各子集的標籤拼法不統一：
- para_breezy 用 fear / surprise，if_48k 用 fearful / surprised。
- 耳語在 para_breezy 是 volume = "whisper"，在 if_48k 是 emotion = "whisper"。
para_tags 型別不同：para_breezy 是 struct，if_48k 是 JSON 字串。

5. 讀取範例

import io, json
import soundfile as sf
from datasets import load_dataset

ds = load_dataset("wilzzzz/TASTE2-SFT-EN", "para_breezy", split="validation", streaming=True)
row = next(iter(ds))

for m in row["message"]:
    print(m["role"], m["timestamp_range"], m["para_tags"], m["text"])
    if m["audio"]:
        wav, sr = sf.read(io.BytesIO(m["audio"]))   # MP3 需要 libsndfile >= 1.1.0

if_48k 的 system prompt 要另外讀 row["system_prompt"]，para_tags 要 json.loads。

只要文字、不想讀音訊時，可以用 pyarrow 只讀巢狀欄位：

import pyarrow.parquet as pq
t = pq.ParquetFile("dialogue/train/stage3_sft_00001.parquet").read(columns=[
    "idx", "meta",
    "message.list.element.role",
    "message.list.element.text",
    "message.list.element.timestamp_range",
])

還原整段對話音檔：建一條全長的空白波形，把每輪 audio 放在 timestamp_range[0] 毫秒的位置；dialogue 有重疊時兩段相加即可。

Downloads last month: 1,505

Total file size:

58.9 GB

You need to agree to share your Hugging Face information to access this dataset

TASTE2 SFT 英文語音對話資料（DATA_SFT_EN）

原始來源

1. 文字在哪裡

2. Parquet 資料結構

2.1 dialogue、control

2.2 para_breezy

2.3 if_48k