skill-extractor (Python)

Extract skills from job postings and resumes. A 32,000-skill gazetteer proposes candidates; a MiniLM + MLP context classifier accepts or rejects each one, so "we value a can-do attitude" doesn't become a skill. Trained on 491K labeled samples, 73% F1 on held-out job postings.

Runs on ONNX Runtime — no torch, no transformers (~60MB of deps, not ~2GB). The MiniLM model (~90MB, or ~23MB quantized) downloads from the Hugging Face Hub on first use and is cached.

pip install skill-extractor

from skill_extractor import extract_skills

text = """Senior Backend Engineer. 5+ years Python or Go, REST APIs with
FastAPI, PostgreSQL, AWS (ECS, Lambda). Docker and Kubernetes required.
Strong communication skills."""

print(extract_skills(text))
# ['aws', 'communication', 'docker', 'ecs', 'fastapi', 'go', 'kubernetes',
#  'lambda', 'postgresql', 'python', 'rest apis']

Lower-level API:

from skill_extractor import SkillExtractor

ex = SkillExtractor(quantized=True)   # 23MB model, ~4x faster, ~same accuracy
ex.extract(text, threshold=0.7)       # stricter
ex.candidates(text)                   # raw gazetteer hits + context windows

Part of the skill-extractor family (Python · JavaScript · Ruby · Rust — identical output, shared parity fixtures). Built by Qarera — see our analysis of 360,000+ job postings for what this pipeline extracts at scale. MIT.