soccernetpy

Python bindings for two text-classification models provided by soccer-rs:

• SOCcerNET classifies job titles and tasks into US SOC 2010 occupations.
• CLIPS classifies product and service descriptions into NAICS 2022 industries.

The computational code is written in Rust and exposed to Python with PyO3. Both the distribution and its import module are named soccernetpy.

Installation

Once published on PyPI:

pip install soccernetpy

For local development, create or activate a virtual environment and run:

maturin develop

To build an installable wheel:

maturin build --release
pip install ../target/wheels/soccernetpy-*.whl

The exact wheel filename contains its supported Python version, operating system, and CPU architecture. A wheel built for one platform may not install on another platform.

The first non-empty classification may download the embedding and classifier model files. Later calls use the cached copies.

Platform support

Prebuilt wheels support both Apple Silicon (aarch64-apple-darwin) and Intel (x86_64-apple-darwin) Macs. Intel wheels bundle the custom ONNX Runtime build from custom-ort-macos-x86-64, because Microsoft no longer publishes Intel macOS binaries.

SOCcerNET

from soccernetpy import soccernet

results = soccernet(
    job_titles=["soccer player", "coach"],
    job_tasks=["play soccer", "coach players"],
    n=3,
)

for job_candidates in results:
    for candidate in job_candidates:
        print(candidate.code, candidate.title, candidate.score)

Signature:

soccernet(
    job_titles,
    job_tasks,
    soc1980=None,
    isco1988=None,
    noc2011=None,
    n=10,
)

job_titles and job_tasks must have equal lengths. Each optional prior-code list must also contain one entry per job. A prior entry may be:

• one code, such as "261";
• multiple codes, such as ["211", "212"]; or
• None when that job has no prior code.

For example:

results = soccernet(
    job_titles=["doctor", "lawyer", "coach"],
    job_tasks=["treat patients", "give legal advice", "train athletes"],
    soc1980=["261", ["211", "212"], None],
    n=5,
)

SOC 1980, ISCO 1988, and NOC 2011 priors are crosswalked into SOC 2010 before inference. Codes not found in a crosswalk do not contribute a prior.

CLIPS

from soccernetpy import clips

results = clips(
    products_services=[
        "Software development",
        "Full-service dental care",
    ],
    sic1987=["7372", "8021"],
    n=3,
)

Signature:

clips(products_services, sic1987=None, n=10)

sic1987 follows the same one-code, multiple-code, or None convention as the SOCcerNET priors. SIC 1987 priors are crosswalked into NAICS 2022.

Results

Both functions return list[list[SoccerResult]]:

• The outer list has one entry per input row and preserves input order.
• Each inner list contains at most n ranked candidates.
• Candidates expose code, title, and score attributes.

best = results[0][0]
print(best.code)
print(best.title)
print(best.score)

n=0 returns an empty candidate list for each input. Empty input lists return an empty outer list without loading a model.

Errors

• Mismatched input lengths raise ValueError.
• A negative n raises OverflowError during Python-to-Rust conversion.
• Model, inference, cache, and crosswalk failures raise RuntimeError.

Development checks

cargo fmt --check
cargo test
maturin develop
python working/test2.py