setfit-example-classifier

Example SetFit classification of job-vacancy postings using a small synthetic dataset.

Project goal

Demonstrate how to fine-tune a Sentence-Transformers encoder with the SetFit framework to distinguish “interesting” tech jobs (senior, remote, well-paid) from everything else.

Quick start

Taskfile handles task management, and uv manages the virtual environment and dependencies automatically (including python executable).

# Train the model and save the best checkpoint under models/
task train

# Evaluate on the held-out test set
task test

# List all available tasks
task --list

Dataset

Sixty short, hand-crafted job-ad snippets labelled 1 = interesting (e.g. “Senior Rust Developer — Remote — $150k”) 0 = uninteresting (e.g. “Call-centre agent — rotating shifts”). See the three CSVs under data/.

Model

• Base encoder: all-MiniLM-L6-v2
• Contrastive fine-tuning + linear classification head via SetFit
• CPU-only by default — change device in main.py if you have a GPU.

License

Released under the MIT License – see LICENSE for details.