BuilderHelm/Kanso-Dictation-Bench-v1

Kanso Dictation Bench v1

A public 1,500-example English text-formatting evaluation from BuilderHelm AI. It tests removal of injected fillers and false restarts, punctuation and capitalization restoration, written numbers, and preservation of already-clean prose.

This is a first-party, synthetic text benchmark. It is not an audio dataset, an ASR evaluation, an independent certification, or a representative sample of real users' dictation. References were frozen before comparative model runs; no evaluated model output is used as a reference.

Composition

Category	Rows	Track
Fillers	200	Formatting
Heavy fillers and restarts	200	Formatting
Numbers, money, dates, and times	250	Formatting
Questions	150	Formatting
Short fragments	100	Formatting
Run-on sentences	150	Formatting
Exclamations	150	Formatting
Already-clean prose	300	Preservation
Total	1,500

Formatting scores use 1,200 rows. Preservation scores use 300 different rows. Do not combine the two tracks into one unlabeled score.

Sources and attribution

• 767 rows adapt assistant-output sentences from the Smol-rewrite test subset of HuggingFaceTB/smoltalk, revision 5feaf2fd3ffca7c237fc38d1861bc30365d48ffa. Created by Hugging Face; Apache-2.0. The repository card identifies Smol-rewrite as an Apache-2.0 component; other SmolTalk components have their own terms and are not used here.
• 383 rows adapt turns from HuggingFaceTB/everyday-conversations-llama3.1-2k, revision 14f543216b9ba42b6b951dc5bd199460d193b162. Created by Hugging Face using Llama-3.1-70B-Instruct; Apache-2.0. Both upstream train_sft and test_sft supply text; the upstream split is recorded on every row.
• 350 rows are original deterministic BuilderHelm AI cases: 250 number/date/time/money examples and 100 short fragments. These original rows and transformation annotations are CC-BY-4.0, attribution BuilderHelm AI, Kanso Dictation Bench v1 (2026).

The source datasets are themselves synthetic. All 1,500 rows are therefore synthetic or synthetic-source adaptations. Source-derived text retains its Apache-2.0 licence. See LICENSE.md and NOTICE, and retain the row-level source and attribution fields when redistributing subsets.

Every source-derived row retains its source repository, revision, split, row index, turn, and sentence-segment index. Original source files and their hashes are recorded in manifest.json. No recordings, private user transcripts, contact lists, or app history are included.

How references and inputs were built

scripts/build_dataset.py fixes seed 20261027, filters source text for basic structural quality, and selects no more than one example from a source conversation/row. Smol-rewrite replies are split into individual sentences before selection; the segment index is retained.

• Fillers: insert a known disfluency into an otherwise unchanged reference sentence, lowercase the spoken input, and remove punctuation.
• Heavy fillers/restarts: prepend a filler stack and a repeated opening span; the reference remains the source sentence.
• Questions/exclamations: remove the source punctuation/casing without inferring a new emotional label.
• Run-ons: remove boundaries from source text containing multiple sentences.
• Facts: derive spoken and written forms from the same explicit numeric value. Target conventions use dollar symbols, digit percentages, h:mm times, Month day dates, and comma-separated integer counts.
• Fragments: preserve a short phrase rather than invent a complete sentence; target capitalization and a terminal full stop are specified by this benchmark's policy.
• Preservation: input and reference are identical; obvious spoken-number normalization cases are filtered from this subset.

These rules express one target rendering. Other punctuation, casing, or numeric styles can be acceptable to a person while scoring below exact match. The dataset does not measure factual correctness of source statements.

Quality and overlap checks

audit.json records exact input/reference/ID/group uniqueness, category counts, source licensing labels, and checks against 13 locally available training/validation/evaluation files from Kanso development. One overlapping public-source question was excluded before the suite was frozen; see exclusions.json. There were no remaining exact overlaps in the inspected files and no reference pairs meeting the audit's 0.85 four-word-shingle Jaccard threshold.

The available local files include earlier training splits and evaluation sets. The current 86,524-pair v22 training split was not present, so it could not be checked directly. Four-word-shingle Jaccard checks (threshold 0.85) supplement normalized exact matching for available files; short fragments use exact matching.

These checks and a stratified agent review are not an independent human annotation study. Related synthetic templates can remain correlated. The stock models' original pretraining data is unavailable, so the dataset must not be described as provably absent from every model's pretraining. The original 500-row historical Kanso suite is not included in these 1,500 rows.

Scoring

scripts/score.py is the versioned reference scorer:

Metric	Definition
WER ↓	Word-aligned substitutions, deletions, and insertions divided by reference words; alignment ignores case. Tables use the mean of per-row rates. Corpus-weighted WER is also exported.
Added content ↓	Word-aligned insertions divided by reference words. This is an insertion proxy, not a factual-hallucination judgment.
Exact match ↑	Case- and punctuation-sensitive equality after whitespace normalization.
Normalized exact ↑	Separately named case-insensitive compatibility metric; not the primary exact-match score.
Punctuation F1 ↑	Precision/recall/F1 of punctuation recovered at aligned positions, averaged across rows.
Capitalization ↑	Matching-case proportion among aligned matched words. Empty-output rows score zero.
Sentence starts ↑	Correct first-word identity and case after reference ., !, or ?.

Tokenization preserves contractions as word tokens and treats other punctuation separately. This intentionally scores written forms literally; it does not apply a semantic number-equivalence normalizer. Abbreviations may introduce extra sentence boundaries. Failure and truncation counts accompany results, and every input must have a prediction.

Reproduce

Apple Silicon and a working MLX Metal runtime are required to generate model predictions. Scoring and structural auditing do not require MLX. The published runs use Python 3.14.7. Start from a clone of a pinned dataset revision and keep its checked-in test bytes; regenerating the dataset is optional.

python3.14 -m venv .venv
.venv/bin/pip install -r requirements.txt
.venv/bin/python -m unittest discover -s scripts -p 'test_*.py'
.venv/bin/python scripts/audit_dataset.py
# Preserve published predictions so the resumable runner creates fresh runs.
mv results results-published
.venv/bin/python scripts/run_models.py
.venv/bin/python scripts/verify_models.py
.venv/bin/python scripts/profile_speed.py
.venv/bin/python scripts/export_results.py
.venv/bin/python scripts/score.py --predictions results/koi-1.0.jsonl --output results/koi-1.0.scores.json

To regenerate source transformations separately, run scripts/build_dataset.py and compare the generated test SHA-256 with the frozen manifest before evaluating. For score-only reproduction, use the checked-in predictions with scripts/score.py; models do not need to be downloaded.

The model manifest pins release revisions, not mutable main branches. Koi 0.9 uses the historical 6-bit bundle at 53bd03b4c69953a8116ecf763bafae017c3ad574; the later repository head contains a corrected base and must not silently replace that historical comparison. Live adapters and each model's required chat template are retained.

All models receive the same formatting system instruction and input text, with greedy decoding, thinking disabled, and a 256-output-token cap. This is a model-only track: ZenVoice's separate formatting rules and content guard are not applied. Run metadata records the actual model-file hashes, runtime versions, machine, and dataset/scorer hashes. Failed/truncated rows remain in the evaluation.

Speed numbers come from two warm model-only passes over a fixed, seeded, stratified 50-example subset (100 observations per model). The first pass follows manifest order; the second reverses it. Load plus first request is the median of two fresh model loads, measured separately; operating-system file caches are not cleared. Per-input timings from the full quality runs are retained separately. The profile records whether repeated outputs match their quality-run outputs. Concurrent system activity is not controlled. Peak memory is MLX Metal allocation, not process RSS or whole-app footprint. GB is decimal. Do not compare these timings directly with unrelated hardware or separately configured app-pipeline runs.

Use and maintenance

Use the frozen test split to evaluate a release. If examples are used to tune a new checkpoint, disclose that exposure and create a new independent test version for claims about unseen performance. Retain the published v1 bytes and revision when comparing old and new results.

@dataset{builderhelm2026kansodictationbench,
  author = {BuilderHelm AI},
  title = {Kanso Dictation Bench v1},
  year = {2026},
  url = {https://huggingface.co/datasets/BuilderHelm/Kanso-Dictation-Bench-v1}
}