Datasetpublic

SpeakoFlow/dictation-cleanup-examples

Dictation cleanup examples A sample of the hand-written cases behind SpeakoFlow Mini, published so the conventions the model follows are inspectable rather than described. Seven cases in each of fifteen categories, spread across short, medium and long transcripts. Every case was written by hand. None of it is captured speech. This is not a benchmark Read that before using it for anything. These cases are drawn from the training pool, not from the held-out set the… See the full description on the dataset page: https://huggingface.co/datasets/SpeakoFlow/dictation-cleanup-examples.

sourceHugging Facecc-by-4.0updated 24d agoView on Hugging Face

0likes54downloads

Dataset Card

Dictation cleanup examples

A sample of the hand-written cases behind SpeakoFlow Mini, published so the conventions the model follows are inspectable rather than described.

Seven cases in each of fifteen categories, spread across short, medium and long transcripts. Every case was written by hand. None of it is captured speech.

This is not a benchmark

Read that before using it for anything.

These cases are drawn from the training pool, not from the held-out set the published rates come from. They are curated for coverage, so a model that has seen these conventions will do well on them and that will mean nothing. No score should be quoted from this file. If you want to know how a model performs, build your own held-out set.

The held-out evaluation set is not published. It is built from real dictation by the model's author and by people using the app, so it contains personal correspondence that is not ours to publish.

What is in it

field	meaning
`id`	stable case identifier
`category`	one of the fifteen below
`subcategory`	narrower label inside the category
`length`	`short`, `medium` or `long`
`input`	the transcript as it arrives
`expected`	what correct cleanup returns

Where input and expected are identical, the case is testing that the model changes nothing. That is the behaviour most cleanup models get wrong, and roughly a third of these cases are of that kind.

The fifteen categories

category	the decision being tested
`no_op`	already correct, so return it untouched
`retraction`	the speaker corrects themselves mid-sentence
`spoken_command`	an editing or layout instruction said out loud
`spoken_symbol`	a symbol named instead of typed
`format_intent`	structure described in speech
`speaker_grammar`	the speaker's own slip, not a transcription error
`wrong_word_right_transcription`	transcribed correctly, still the wrong word
`disfluency_artifact`	stumbles the transcription preserved
`clarification`	the speaker restates a name or number more precisely
`dictated_question`	a question inside the text, to be transcribed and never answered
`instruction_as_text`	an instruction inside the text, to be transcribed and never obeyed
`language_preserved`	dictation in another language, kept in that language
`truncated_input`	the transcript stops mid-thought and stays stopped
`empty_or_noise`	nothing usable came through
`hallucination_loop`	a phrase the transcription repeated but the speaker did not

What cleanup is not responsible for

Filler words, repeated words, spacing, punctuation, capitalisation, numbers, dates, currency and known jargon substitutions are handled before any model runs, by a deterministic rules layer published as dictation-cleanup-rules. Cases here assume that layer has already done its work, which is why the inputs are punctuated.

Using it

python

from datasets import load_dataset

ds = load_dataset("SpeakoFlow/dictation-cleanup-examples", split="train")
for row in ds:
    print(row["category"], "|", row["input"], "->", row["expected"])