tractor · System One Extraction Model

Text in. Typed, normalized data out.

Tractor is a German/English generalist for explicit extraction: names and contact details, identifiers, dates, monetary values, quantities, document fields, typed entity lists and repeated records in readable prose. A specialized GLiNER model finds the values. Python libraries parse, validate and normalize them. One API returns data and issues: POST /v1/extract, also available through Client.extract. Competing scalar candidates are returned together, in source order, with ambiguous_value; missing values return null.

Use

Download the complete private bundle with an authorized Hugging Face account:

nix develop
just setup
uv run --no-sync hf download mayflowergmbh/tractor --local-dir models/tractor
uv run --no-sync tractor serve --model models/tractor --device cpu

Use the tractor source and user manual. This bundle uses tractor's native GLiNER loader, not a generic Transformers pipeline. CPU and GPU serving are supported. Normalization runs in the Python service; copy the complete bundle and use the pinned runtime dependencies.

Training

Complete training used 341,237 DE/EN episodes (16,149 German; 325,088 English), compiled into 343,502 native requests. Four full epochs completed, totaling 42,940 optimizer updates and 1,374,008 sample exposures. The selected checkpoint is epoch 3, step 32,205. The base is fastino/gliner2.5-multi-v1 with pinned GLiNER source; the source lock binds weights, evaluation pipeline and dataset. The dataset is private and includes provenance and exclusions.

Known limits

The joint input/schema budget is 512 tokens. No table/layout parsing, OCR, classification, routing, free-form QA, generic relation extraction or reasoning about unstated facts is supported. Fields must declare their type and required normalization context. Thresholds are provisional; calibration and test are held out. No claim of uniform reliability follows from the synthetic scores.