Skip to content

Examples

These examples are intended as inspiration for data handling and workflow variations. They are not installed with the pip package; see gradiend/examples on GitHub to download a file or read to get inspired.

Quick start

  • start_workflow.py — Minimal start-to-end workflow with TextPredictionDataCreator. Uses 75+ artificial sentences (3SG/3PL), creates data on the fly, trains, and evaluates. Matches docs/start.md.

Training workflows

  • gender_de.py — Gender bias in German (minimal workflow, single pair).
  • gender_de_detailed.py — German gender workflow for various target class combinations with pruning, encoder plots, top-k overlap heatmap, and training convergence.
  • gender_de_decoder_only.py — Decoder-only model with optional MLM head (DecoderModelWithMLMHead).
  • gender_en.py — Gender bias in English with name augmentation and GENTypes-based decoder metrics (BPI, FPI, MPI).
  • race_religion.py — Race and religion bias (multi-class, multiple bias types in a loop).
  • english_pronouns.py — English pronouns (3SG vs 3PL); loads data from data_creation_pronouns output. Optional class_merge_map for number/person (e.g. singular vs plural). See also the english_pronouns.ipynb notebook for a step-by-step run (data creation from Wikipedia → training → evaluation).

Data creation

  • data_creator_demo.py — Build training and neutral data with TextPredictionDataCreator (German articles der/die/das, syncretism handling).
  • data_creation_pronouns.py — Create English pronoun data (1SG, 1PL, 2, 3SG, 3PL) from Wikipedia via HuggingFace; used by english_pronouns.py.

Notebooks (interactive)

Interactive Jupyter notebooks for step-by-step workflow discovery:

  • gender_de_detailed.ipynb — Start with a single gender-case pair, explore each step, then optionally loop over configs and compare with top-k overlap heatmaps.
  • english_pronouns.ipynb — Data creation from Wikipedia → training 3SG vs 3PL → evaluation, with optional class_merge_map for singular vs plural.