flowcorp-ch/BudgieScribe-eval

BudgieScribe held-out evaluation sets

The synthetic held-out sets every BudgieScribe build is scored on. Deterministic generators (github.com/alexxxcoelho/budgie-scribe, scribe/generateurs/gen_*.py, fixed seeds) start from the value, so the expected output is known to the character and scoring is exact string comparison — no judge. Each generator writes 8 % of what it produces into a held-out slice, by family, never seen in training. These files are that slice.

Language	File	Axis	Cases
fr	fr/heldout_forme.jsonl	Formatting blocks (lists, emails, addresses)	635
fr	fr/heldout_correction.jsonl	Speaker self-corrections	476
fr	fr/heldout_itn.jsonl	Inverse text normalization (numbers, amounts, dates, times, phones)	797
fr	fr/heldout_compo.jsonl	Composition: 2–3 adjacent numbers in one sentence	319
fr	fr/heldout_cor_itn.jsonl	Self-correction × number in one sentence	238
en	en/heldout_forme.jsonl	Formatting blocks	634
en	en/heldout_correction.jsonl	Speaker self-corrections	474
en	en/heldout_itn.jsonl	Inverse text normalization	798
en	en/heldout_compo.jsonl	Composition	319
en	en/heldout_cor_itn.jsonl	Self-correction × number	237

Row format

One JSON object per line:

{"id": "itn-00005", "file": "itn-telephone", "source": "itn", "lang": "fr",
 "held_out": true,
 "styling": "semi-formal", "structure": "prose", "context": "general",
 "control": "[Styling: semi-formal] [Structure: prose] [Context: general] [Lang: fr]",
 "dirty": "tu peux m'appeler au zéro six douze trente-quatre cinquante-six soixante-dix-huit",
 "clean": "Tu peux m'appeler au 06 12 34 56 78."}

dirty is what an ASR engine produces; clean is the expected output; control is the line the model receives above the transcript (see FORMAT.md). file carries the family (itn-telephone, frm-liste, …) the benches report by.

Scoring a build

hf download flowcorp-ch/BudgieScribe-eval --repo-type dataset --local-dir eval
export SCRIBE_LANG=fr SCRIBE_TRAVAIL=$PWD/eval
python scribe/bancs/eval_synth.py <checkpoint-dir | base | base:mini> eval/fr/heldout_forme.jsonl

or, on Hugging Face Jobs, hf/jobs/eval.py --model <repo> --lang fr from the code repository, which runs every file of the language. Reported per axis: exact matches, and for numbers wrong value separately from wrong formatting.

Personal data

Nothing here comes from a recording or a person. Every string is generated from templates and random values. Before publication the sets were passed through the project's PII scanner (scribe/pipeline/pii_scan.py); the e-mail domains the generators use (gmail.com, outlook.fr, free.fr, …) were replaced on both sides of each pair by RFC 2606 reserved domains (example.com, example.org, example.net) so that no plausible real address is distributed. Phone numbers are random digit strings in the national formats; any match with a real number is coincidental. The published model tables were measured on the sets before the domain swap; the swap changes the spoken domain only, and a re-run is expected to differ by at most a few cases on the e-mail families.

License

Apache 2.0, like the generators that produce them.