README.md

---
license: apache-2.0
language:
- en
task_categories:
- audio-to-audio
tags:
- full-duplex
- spoken-dialogue
- turn-taking
- instruction-following
- synthetic
---

# FD_data_v2 — synthetic full-duplex spoken conversations with instruction-conditioned turn-taking

Fully synthetic two-speaker conversations where the assistant's turn-taking
behaviour is conditioned on a spoken instruction, with construction-time
ground-truth timestamps. Voices are synthesized with
[Kokoro-82M](https://huggingface.co/hexgrad/Kokoro-82M) (Apache-2.0);
transcripts are synthetic and contain no personal data.

## What a row is

One row is one VARIANT. Rows sharing a `base_conv_id` share a
sample-identical user channel and differ only in what the assistant does, so
they can be compared directly (and used as preference pairs).

- `audio` — stereo 24 kHz WAV, **channel 0 = user, channel 1 = assistant**.
  Overlap (backchannels, interruptions) is real simultaneous audio across the
  two channels, not concatenation.
- `instruction` — what the user asked for, in speech. Empty for the
  un-instructed baseline variant.
- `program` — JSON, present when the instruction was sampled from a grammar
  rather than taken from a fixed policy list. Ordered rules of
  `(condition, action, scope, at_most)` over observable atoms, with
  and/or/not, ordinals ("the third time I pause"), scopes
  ("until I say otherwise") and rule precedence.
- `decisions` — JSON, the action the interpreter computed at every decision
  point: `{moment, turn, action, rule}`. **The labels are computed by running
  `program` over the conversation, not asserted**, and re-running it
  reproduces them exactly.
- `turns` — JSON: speaker, text, absolute start/end, word-level timestamps
  (MMS forced alignment), plus `cut` and the `unspoken` tail when a turn was
  interrupted. **The full reference text is retained** — it is what the speaker
  meant to say — so on a cut turn some of `text` and some of `words` is not in
  the audio. Every word therefore carries `spoken`: filter on it before using
  the timestamps for anything acoustic. A word the cut lands inside counts as
  unspoken.
- `events` — JSON, the timeline: `pause`, `assistant_backchannel`,
  `assistant_interrupt`, `user_backchannel`, `user_interrupt`, `tool_call`,
  `tool_result`. Timestamps are constructed, not estimated.
- `realization` — JSON, how the label was carried in audio: `yield_decision`
  (did the assistant give up the floor when cut into) and `response_latency`
  (how fast it answered, measured from when the floor came free).
- `tool` — JSON for tool rows: the call goes out on a text stream WHILE the
  assistant speaks a hold phrase. Cancelled calls carry a `tool_call` and
  deliberately no `tool_result`.
- `verify_pass` / `verify_checks` — deterministic Track-A verdict. The event
  structure is re-derived from the assembled audio with Silero VAD and checked
  against the script, including that the label's audible consequence is
  actually present.

## Actions

`listen`, `backchannel`, `interrupt`, `continue`, `acknowledge` — the taxonomy
of Instruct-FD (arXiv 2607.20460), so scores stay comparable.

## Honest limits

- Timing bands (gaps, reaction latencies) are PLACEHOLDERS, not yet calibrated
  against a real spoken corpus.
- Kokoro voices only; no expressive-TTS subset in this batch.
- Verification is deterministic (structure, timing, whether the label is
  audible). It is not a content-quality review.
- Batch scale, not training scale.

## 字段取值说明（中文）

### `kind` — 场景大类

| 取值 | 含义 |
| --- | --- |
| `proactive` | 用户持有发言权，助手判断要不要以及何时介入（打断 / 附和 / 继续听） |
| `responsive` | 助手持有发言权，用户压着它说话，助手决定让不让出发言权 |
| `tool` | 助手嘴上说 hold 语的同时，文本流并行发出工具调用 |

### `policy` — 这一行的指令类型

`none` 是同组的无指令对照行。`prog0`/`prog1` 只是编号，真实语义在 `program`
字段里的规则，渲染成口语后放在 `instruction`。其余是固定策略名，取自
Instruct-FD（arXiv 2607.20460）的动作集合，便于与已发表结果对齐。

| 取值 | 含义 |
| --- | --- |
| `none` | 无指令基线。同组的对照行，用来看模型在没有指令时的默认行为 |
| `listen` | 安静听着，不打断也不附和，等用户明确说完再接话 |
| `backchannel` | 在用户停顿或寻求确认时给简短回应（嗯、好的），不夺发言权 |
| `interrupt` | 用户说错时立刻打断纠正，不等他说完 |
| `continue` | 自己正在说话时，用户给出简短附和，视为支持，说完不让 |
| `acknowledge` | 自己正在说话时，用户插入纠正或新约束，立刻停下、确认并调整 |
| `prog0` | 采样出来的指令程序（第 1 条）。真实语义在 `program` 字段，不在这个名字里 |
| `prog1` | 采样出来的指令程序（第 2 条），与 prog0 在同一段对话上要求不同的行为 |

### `scenario` — 具体情境

标【诱饵】的场景里，出现的东西**看着像**该介入的信号，但正确做法是不动。

| 取值 | 情境 |
| --- | --- |
| `word_retrieval_assist` | 用户想不起某个词，绕着这个概念描述——要不要替他补上 |
| `factual_misinformation` | 用户很自信地说错了一个日常事实，并开始据此做计划 |
| `sequential_info_capture` | 用户在念结构化信息（地址、编号），需要逐条确认收到 |
| `self_contradiction` | 用户这句话和自己前面说过的直接矛盾 |
| `hesitation_prompt` | 用户在排练发言，卡住并出现长时间迟疑 |
| `cognitive_pause` | 用户句子说到一半陷入沉默思考，没有任何口头信号，然后继续 |
| `third_party_aside` | 【诱饵】用户扭头对房间里另一个人说话——听着像在问助手，但不是 |
| `hypothetical_framing` | 【诱饵】用户明说那是别人的看法或一个假设，不是他自己信的 |
| `emotional_escalation` | 用户对一直修不好的事情明显烦躁起来 |
| `user_attention_check` | 用户中途确认助手还跟得上（「你懂我意思吧」） |
| `emotional_disclosure` | 用户讲一段带情绪的私人经历，中途留下一个沉重的停顿 |
| `safety_correction` | 用户轻描淡写地说了一个有安全风险的错误认知，并打算照做 |
| `user_filler_pause` | 用户边想边说，带着「呃」「嗯」和规划性的停顿 |
| `clause_boundary_tracking` | 用户一口气讲一条多段推理，子句边界清楚但完全没有停顿或确认 |
| `mid_turn_direct_question` | 用户句中真问了助手一个能回答的问题（不是反问），并且不等回答继续说 |
| `user_self_correction_midstream` | 【诱饵】用户正要说错一个数字或事实，自己中途察觉并当场改正 |
| `hearing_check` | 用户插话说没听清最后一段，要求重说 |
| `scope_narrowing` | 助手在泛泛回答，用户插话把范围收窄到某一点 |
| `topic_redirect` | 助手在答一件事，用户插话转向另一件更要紧的事 |
| `urgent_stop` | 助手正在执行，用户紧急叫停 |
| `assistant_self_repair` | 助手自己说错了一处，用户插话指出 |
| `short_stop_repair` | 用户用「等一下」「不是」这类短促信号打断 |
| `user_acknowledgment` | 助手解释时用户短暂重叠一句附和，表示在听 |
| `user_continuer` | 助手解释时用户重叠一个「接着说」类的信号 |
| `add_constraint` | 助手正在讲方案，用户插进来加一条新约束，方案要跟着改 |
| `tool_info_request` | 用户问的问题只能靠查询工具回答，助手说 hold 语的同时发出调用 |
| `tool_action_request` | 用户要求执行一个动作，助手说 hold 语的同时发出调用 |
| `tool_followup_constraint` | 用户先提要求，随后补一个额外约束，调用要反映补充后的要求 |
| `tool_barge_in_during_hold` | 助手正说 hold 语时用户插话改需求，调用必须反映改动 |
| `tool_cancel_during_hold` | 助手正说 hold 语时用户取消——调用已在途中，且**不能**报结果 |