tractor · System One Extraction Model

Text in. Structured data out. Ready for your application.

tractor is Mayflower’s System One Extraction Model: a specialized German/English model and API for turning readable text into typed fields, separate records and normalized values. Describe the fields you need. Get JSON your code can use.

Extract names, contact details, identifiers, dates, amounts and quantities from business documents, messages and other prose. Learned extraction combines with programmatic normalization using established Python libraries. Run the complete pipeline on your own infrastructure, on CPU or GPU.

User manual · Get started · API · Benchmarks · Model card

Put extraction on the fast path

Your application needs	tractor provides
Fields from messy text	Source-span extraction with field descriptions in plain language
Consistent typed values	Library-backed parsing, validation and deterministic normalization
Repeated records in prose	Anchored records with fields assigned to each explicitly described entry
Explicit outcomes	Missing-value, normalization and ambiguity issues
Control over deployment	Local, offline serving with CPU or GPU support
A small integration surface	One HTTP endpoint and a lightweight Python SDK

A single encoder pass finds the relevant spans; deterministic code parses and normalizes the results. Keep repeated extraction local, with one public API and explicit handling of missing or ambiguous values.

Specialized extraction, measured

99.09% field F1 on synthetic single-object Dev requests. 90.30% on fields in repeated prose records.

Version 2026.10.21 uses the selected checkpoint from complete training on 341,237 German/English episodes. Evaluation covers 4,483 Dev documents through the full API pipeline. Complete-record F1 is 57.61%: strong field extraction does not mean every record is complete. Real human-annotated Dev field F1 ranges from 71.03% for contracts to 92.67% for invoices in the measured sets.

Dev selected the checkpoint; these are not independent test results. Strict quality and regression gates are not all met. See results, evidence and known limitations. No latency or Haiku/Luna comparison is claimed for these weights.

From source text to application data

A message says “Die Breite beträgt 12,5 cm.” Your system needs millimeters. tractor returns 125 under your field name. Other declared field types select their own parsers and normalizers. Repeated records preserve their field assignments; missing or ambiguous values produce structured issues.

Use it for document metadata, CRM fields, explicitly named contract parties, contact details, product attributes and measurements. Inputs and schemas share a 512-token budget. Validate the complete pipeline on your own workload.

Start the API

Use the Nix flake development environment. Authenticate with hf auth login using an account with access to the private model repository, then download the bundle and start the API:

nix develop
just setup
uv run --no-sync hf download mayflowergmbh/tractor --revision v2026.10.21 --local-dir models/tractor
tractor serve --model models/tractor --device cuda

Use --device cpu for CPU inference. Serving loads local files offline and does not download or train a model. GET /readyz reports readiness; interactive API documentation is available at http://127.0.0.1:8000/docs.

curl http://127.0.0.1:8000/v1/extract \
  -H 'Content-Type: application/json' \
  -d '{"text":"Die Breite beträgt 12,5 cm.","locale":"de-DE","fields":{"width_mm":{"type":"quantity","description":"Breite","unit":"mm"}}}'

{"data":{"width_mm":125},"issues":[]}

Python SDK

from tractor import Client, Quantity, Text

with Client("http://127.0.0.1:8000") as client:
    result = client.extract(
        "Alpha: Breite 10 cm. Beta: Breite 12 cm.",
        {"product": Text("Produktbezeichnung"),
         "width_mm": Quantity("Breite", unit="mm")},
        anchor="product",
        many=True,
    )
    print(result.data)

The result contains separate Alpha and Beta records with widths 100 and 120 mm. The public response contains only data and issues. Missing values produce null; competing scalar candidates are returned together with ambiguous_value. The SDK imports no ML libraries; just setup-sdk installs its development dependencies without the model stack.

Evaluation and limits

The selected model improves generalist coverage, with remaining weaknesses in complete records, restaurant cuisine, product prices and some company/contract fields. 12 regression checks remain failed. Real evidence is missing for many domains, and incomplete references are reported as unscored. Calibration and test remain held out. See the full evaluation and model card.

Development

nix develop
just setup
just check
just check-model models/tractor cpu

Python dependencies and native libraries are pinned in uv.lock and flake.lock. The CLI also provides data generation, validation, training, evaluation and benchmark commands; run tractor --help. See architecture and upstream compatibility.

Container

docker build -t tractor:local .
docker run --rm --user "$(id -u):$(id -g)" -p 127.0.0.1:8000:8000 \
  -v "$PWD/models/tractor:/model:ro" tractor:local

The container defaults to CPU. GPU serving requires NVIDIA Container Toolkit; add --gpus all and append --model /model --device cuda --host 0.0.0.0.

Source: mayflower/tractor. Model: mayflowergmbh/tractor. Both repositories are private during release preparation. Release licensing is not yet selected.