Borcherding/voice-decisions-v4

Voice Decisions v4

Decision rows for the three calls voice mode makes while a person talks to the assistant in Unsloth Desktop and Unsloth Studio: have they finished, what should the assistant say while it thinks, and what to do when they talk over it.

Rows use the Unsloth Decision format (state, questions, gold), so Unsloth Studio's decision trainer fine-tunes on them directly. One model answers all three questions.

Questions

question	type	asked when	answers
turn_done	yes/no	the person pauses	have they finished, so the assistant should answer now?
filler	pick one of 12	a request was just sent	which kind of short line to say aloud while the model thinks, or none
interject	pick one of 4	the person speaks while the assistant is still thinking	amend, replace, queue or ignore

Interject classes

label	what the app does	examples, said over "write me a history of Rome"
amend	keeps the request, adds these words to it, starts over	"and add some dates", "make it shorter", "hold on", "wait start over i worded that badly"
replace	drops the request, sends only these words (or just stops)	"never mind, what's the weather", "stop", "ignore that", "never mind found it"
queue	keeps the request, answers this after the reply	"also, what time is it in Tokyo"
ignore	nothing to act on	"uh huh", "can you hear me", "hold on mom", "not you, I'm on the phone"

Cue words don't decide the class; what follows them does. A correction to one detail ("no wait, the eastern empire") is amend, however strong the "no". A bare "hold on" said to the assistant is amend; said to someone else it's ignore. When amend and replace are hard to tell apart, the label is amend. The full rubric with boundary pairs is in RUBRIC.md.

Filler classes

minimal, task, code, explain, plan, creative, compare, transform, advice, opinion, correction, and none (say nothing: greetings, yes/no, control commands, quick lookups, short answers to the assistant's own question). Each class maps to a bank of short, neutral phrases in Unsloth Desktop, such as "Let me look at that code." for code. A state starting with (the assistant just asked them a question) marks an answer to the assistant.

Schema

field	type	description
state	string	what the model sees: the words so far (turn_done, filler), or the request plus the words spoken over the thinking and when (interject)
questions	JSON string	one question with its type, instructions and, for choice questions, the class definitions
gold	JSON string	the answer, e.g. {"interject": {"label": "amend"}}; turn_done uses "true" / "false"
origin	string	<task>/<split>/<source>; the task prefix for turn_done rows is judge

{
  "state": "Request: plan a trip to the grand canyon\nSpoken over the thinking (4 seconds into thinking): with a toddler",
  "questions": "{\"interject\": {\"type\": \"choice\", \"instructions\": \"...\", \"criteria\": {\"amend\": \"...\", \"replace\": \"...\", \"queue\": \"...\", \"ignore\": \"...\"}}}",
  "gold": "{\"interject\": {\"label\": \"amend\"}}",
  "origin": "interject/train/claude-interject-b"
}

Splits

split	turn_done	filler	interject	rows
train	13,358	12,444	11,712	37,514
eval	1,462	1,396	1,590	4,448
eval_real	716	580	22	1,318

Every source row appears twice: once lowercase with no punctuation, as speech recognition often writes it, and once cased and punctuated, so casing carries no signal about where a row came from.

• eval is 12% of the synthetic rows, stratified by task and label. Every row built on one text (a finished ask, its cut prefixes, its filler row and the interject pairs on it) stays on one side.
• eval_real is 20% of the real rows, stratified the same way, and never in train. Synthetic rows sharing half or more of their words with an eval_real text were dropped, so paraphrases don't leak it into train.
• No normalized duplicate appears in both train and an eval split.

Composition

source	what it is	train	eval	eval_real
claude-*	written by Claude subagents to a shared spec: turn ends and mid-thought pauses, filler requests, interject pairs, and the hard interject kinds (cancels, "ignore that", "found it", rephrase, talk aimed at someone else)	18,610	2,476
augment-*	paraphrases of real messages, written by Claude subagents	6,104	852
colab-clean	generated with a 12B model in Unsloth Studio on a Colab GPU, then cleaned: templated and near-duplicate rows dropped, labels checked and relabelled	5,302	794
real	the dataset author's own messages to a voice assistant and a coding agent, redacted	5,102		1,318
balance-*	rows written to lift the smallest filler classes, "give me two minutes" interjects, and turn_done prefixes cut at a pause word ("can you write me a")	2,396	326
total		37,514	4,448	1,318

Counts are built rows, after the casing copies.

Class balance

Train is spread across classes on purpose: none is capped near 20% of filler, every filler class has at least 380 source rows, turn_done is about one third complete, and the four interject classes sit between 23% and 26%. eval_real keeps the author's real mix, which is complete-heavy and none-heavy, with only 11 interject pairs.

Utterances stay at 60 words or fewer. Real messages run longer in the tail than synthetic ones, which is part of what eval_real checks.

Personal data and redaction

The real rows come from the author's own conversations. Before release:

• names, numbers, machines, repositories and tool names were swapped for generic words;
• every row was scanned by script for emails, URLs, file paths, long digit runs, issue and PR numbers, file names and keys, plus a crude-word list;
• work-specific details and crude words were removed from real rows and their paraphrases (90 rows).

A rescan of all 43,280 rows and the source file for this release found no emails, URLs, file paths, @handles, key-like strings or runs of six or more digits.

How it was made

Every row was checked by script, then read by hand where it mattered:

• Hard rules on every row: no PII patterns, English only, at most 60 words.
• Read in full: about 580 interject rows with "ignore that", "found it", "start over", "rephrase", "hold on" or "give me a minute", against the rubric (3 relabelled).
• Read on a sample: 400 random rows, then about 800 rows a label heuristic flagged, such as a finished turn ending on "the" or a none filler over twelve words. Sample error rate was about 1%; 16 rows were dropped and 6 relabelled.
• Deduplicated on normalized text across files (672 rows). Same text with different labels kept the real row's label or was dropped (99 rows).
• Interject requests arrive as two glued chat bubbles a quarter of the time, the way the app sends a request said in two breaths, and interject states say how many seconds into thinking the person spoke.

Use

from datasets import load_dataset

ds = load_dataset("Borcherding/voice-decisions-v4")
print(ds["eval_real"][0]["state"])

Intended for training and evaluating small decision models for voice turn-taking in Unsloth Desktop and Unsloth Studio. It is not a speech dataset (there is no audio), and its labels encode one app's behaviour, so it is not a general benchmark for turn-taking.

To train in Unsloth Studio, put the train and eval files under ~/.unsloth/studio/assets/datasets/ and start a decision run on them.

Files

• voice_decisions_v4_{train,eval,eval_real}.jsonl: the dataset.
• stats.json: row and label counts per split, task and source.
• source/rows_v4.jsonl: the source rows before the Decision format (split, task, text fields, label, src, kind).
• RUBRIC.md: the interject labelling rubric.
• charts/: the figures above, built from source/rows_v4.jsonl.

Limitations

• Mostly written, not spoken. About 85% of rows are written by Claude subagents or a 12B model in the style of speech recognition output. eval_real is the check on that gap.
• One speaker's real data. The real rows are one person's messages, much of it about software work, so the real style is narrow.
• Thin classes. eval_real has 11 interject pairs (7 amend, 2 ignore, 1 replace, 1 queue), 2 plan rows and no compare rows: read those scores as spot checks. Real-speech queue and varied ignore backchannels are rare across the whole set.
• Known gaps in coverage, seen in use with models trained on v4:
  • questions or complaints aimed at the assistant mid-thought ("why did you stop?", "can I interrupt you?") should be replace, and have few rows;
  • garbled speech recognition aimed at the assistant tends to read as ignore;
  • direction changes like "wait no, make it about Greece instead" should be replace, and are under-represented next to one-detail amend corrections;
  • quick-fact lookups ("what's the capital of France?") should be filler none.
• Not human-reviewed row by row. Labels come from Claude subagents and a 12B model working from the rubric, with the audit above. Filler classes overlap at the edges (a short status question could be none or minimal).
• English only.

Trained models

• Borcherding/voice-decisions-laya-v4: full fine-tune of Laya.
• Borcherding/voice-decisions-qwen-v4: LoRA on Qwen3.5-0.8B.

This dataset was built to train decision models with Unsloth.