arjun10g/na-tech-jobs-skills-v1

na-tech-jobs skills extractor — v1

Zero-shot skill extraction wrapper that pulls programming languages, frameworks, cloud platforms, databases, and ML libraries from a job description, then normalizes outputs into a fixed canonical taxonomy of ~70 skill names so downstream consumers see consistent labels (py / python / Python all collapse to Python).

Per CLAUDE.md §7 the v1 design is intentionally training-free: rather than fine-tuning, we delegate the extraction to numind/NuExtract-tiny-v1.5 (a 0.5B-param structured-extraction model) and apply our own taxonomy post-processor. The wrapper code lives in the project repo (models/skills/predict.py).

What this repo contains

• The skill taxonomy and alias map.
• A pointer to NuExtract for the actual model weights.
• The wrapper class SkillExtractor.

There are no fine-tuned weights to ship — the underlying model is loaded directly from numind/NuExtract-tiny-v1.5.

Usage

from models.skills.predict import SkillExtractor

extractor = SkillExtractor()
skills = extractor.extract(
    description_md="We use PyTorch, AWS SageMaker, and Kubernetes…",
    title="Senior ML Engineer",
)
# → ["PyTorch", "AWS", "Kubernetes"]

Status (v1)

• Available: ad-hoc per-job calls in app/ and notebooks.
• Not yet batch-applied to the curated table because NuExtract on Apple MPS clocks 1.7 sec/row → ~6 hours for 12k rows. v1.1 will run this batch on an A10G HF Job (30 min).
• v1 enriched parquet exposes extracted_skills_v1 as an empty list for every row; consumers should treat it as "not yet computed."

Citation

Ghumman, A. (2026). na-tech-jobs skills extractor v1. https://huggingface.co/arjun10g/na-tech-jobs-skills-v1