Buttermilk03/scriber-lfm2.5-350m-polishing-de-training-v1

Scriber LFM2.5 German STT post-processing data

This repository contains the exact 2,000 German source/target pairs used to train the final Scriber LFM2.5 350M local post-processing model.

The matching model is `Buttermilk03/scriber-lfm2.5-350m-polishing-de-qad-v1`. The complete production recipe and the lessons that determined it are in `TRAINING.md`; machine-readable settings are in `training_recipe.json`.

Data

Each JSONL row contains:

• —source: flat German speech-to-text input with realistic dictation errors.
• —target: the intended language, typography, number, unit, paragraph, and list correction without changing content.
• —id, source_index, topic, error_tags, and list metadata for auditing.

The source Word document contained 2,000 tables with one source and one target each. Native bullet and numbered-list structure was preserved during extraction.

All 2,000 IDs and all 2,000 source/target pairs are unique; no source or target is empty.

Split design

The source examples were partly ordered by error type, so a numeric 80/10/10 cut would have produced biased splits. Seed 3502026 first reserved 200 final rows by complete canonical topic family. Ten thousand seeded candidates were then searched for a feature-balanced 1,600/200 train/validation assignment. Membership and row order were both randomized.

The split files preserve the research lineage. For the final production SFT, all three files were recombined and every one of the 2,000 original pairs was trained exactly once before deterministic short-example augmentation.

dataset_stats.json records the complete split audit. Its source path was reduced to a filename before publication; source byte size and SHA-256 remain unchanged.

Scope and limitations

The corpus is specialized for German dictated letters and structured business prose. Addresses, organizations, amounts, dates, units, and legal-style references are deliberately present as editing challenges. Pattern checks found no actual e-mail address, phone number, IBAN, or API token in the 2,000 pairs, but users should still review the data for their own requirements.

The data does not by itself make a safe editor. Scriber places a host-side preservation gate around model output and returns the source unchanged when protected values or structure are not preserved.

Attribution

The final weights were trained with Praxist by Sapient Intelligence. This dataset card describes the Scriber data lineage; model and training-tool licenses remain separate and are documented in the model repository.