FD_data_v3 — 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 (Apache-2.0); transcripts are synthetic and contain no personal data.

New versus FD_data_v2: 2.8x the rows, program instructions now cover assistant-held moments (continue / acknowledge are computed labels, not only catalog policies), three-way contrast groups (none / two programs demanding opposite reactions to the same interjection), a split column with an eval split whose voices (the British-accent pool) never appear in train, and composed exception clauses that fire ("stay quiet — but if I'm about to do something unsafe, cut in").

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.
• instruction — what the user asked for, in speech. Empty for the un-instructed baseline variant.
• instruction_kind — catalog (fixed policy taxonomy, many phrasings, optionally composed with an exception clause) or program (sampled from a grammar; see program).
• program — JSON, for program rows: ordered rules of (condition, action, scope, at_most) over observable atoms, plus separate fallback actions for user-held and assistant-held moments (the two floors share only listen, so a single fallback would collapse the label mix).
• decisions — JSON, the action the interpreter computed at every decision point. Labels are computed by running program over the conversation, not asserted; re-running reproduces them exactly.
• turns — JSON: speaker, text, absolute start/end, word-level timestamps, cut and the unspoken tail when a turn was interrupted. Every word carries spoken: filter on it before using timestamps acoustically.
• events — JSON, the timeline: pause, assistant_backchannel, assistant_interrupt, user_backchannel, user_interrupt, tool_call, tool_result. Timestamps are constructed, not estimated.
• realization — JSON: yield_decision (did the assistant give up the floor when cut into) and response_latency (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; barge-in rows emit the call only after the user's amendment finishes, and the arguments reflect it.
• split — train / dev / eval, assigned by topic. The eval split renders with a held-out voice pool (British-accent Kokoro voices) that train never sees.
• 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. Every published row passes; whole groups containing any failing row were dropped (listed in the release notes below).

Actions

listen, backchannel, interrupt, continue, acknowledge — the taxonomy of Instruct-FD (arXiv 2607.20460).

Acceptance measured on this batch

• Label replay: 0 mismatches (stored decisions == re-running the program).
• Instruction-blind leakage (leave-one-group-out oracle): 0.480, below the majority-class baseline 0.501; discrimination 86/86 program groups.
• Track-A: every published row passes (817/823 in the raw batch; the 6 failures and their sibling rows are excluded).

Honest limits

• Timing bands (gaps, reaction latencies) are PLACEHOLDERS, not calibrated against a real spoken corpus.
• Kokoro voices only. The eval split's held-out VOICES are a weaker guarantee than a held-out engine.
• Verification is deterministic (structure, timing, label audibility) plus an ASR fidelity floor. It is not a content-quality review.