Detoxify Versions Save

What's Changed

• [pre-commit.ci] pre-commit suggestions by @pre-commit-ci in https://github.com/unitaryai/detoxify/pull/78
• Limit pytorch lightning version to prevent breaking changes by @dcferreira in https://github.com/unitaryai/detoxify/pull/83
• Update mac version to latest in github workflow by @laurahanu in https://github.com/unitaryai/detoxify/pull/89
• [pre-commit.ci] pre-commit suggestions by @pre-commit-ci in https://github.com/unitaryai/detoxify/pull/88
• Bump transformers from 4.22.1 to 4.30.0 by @dependabot in https://github.com/unitaryai/detoxify/pull/85
• fix for transformers 4.35.2 by @Vela-zz in https://github.com/unitaryai/detoxify/pull/98
• Workflow to update pypi by @laurahanu in https://github.com/unitaryai/detoxify/pull/102

New Contributors

• @dcferreira made their first contribution in https://github.com/unitaryai/detoxify/pull/83
• @dependabot made their first contribution in https://github.com/unitaryai/detoxify/pull/85
• @Vela-zz made their first contribution in https://github.com/unitaryai/detoxify/pull/98

Full Changelog: https://github.com/unitaryai/detoxify/compare/v0.5.1...v0.5.2

What's Changed

• Run tests workflow on a schedule by @jamt9000 in https://github.com/unitaryai/detoxify/pull/51
• [pre-commit.ci] pre-commit suggestions by @pre-commit-ci in https://github.com/unitaryai/detoxify/pull/47
• Map loaded checkpoint to specified device by @gregpriday in https://github.com/unitaryai/detoxify/pull/54
• [pre-commit.ci] pre-commit suggestions by @pre-commit-ci in https://github.com/unitaryai/detoxify/pull/55
• Parameters in evaluate.py by @anitavero in https://github.com/unitaryai/detoxify/pull/57
• [pre-commit.ci] pre-commit suggestions by @pre-commit-ci in https://github.com/unitaryai/detoxify/pull/67
• fix worker numbers type by @omidforoqi in https://github.com/unitaryai/detoxify/pull/68
• Remove --use-feature=2020-resolver pip flag by @jamt9000 in https://github.com/unitaryai/detoxify/pull/73
• Pin transformers version to 4.22.1 by @laurahanu in https://github.com/unitaryai/detoxify/pull/76

New Contributors

• @pre-commit-ci made their first contribution in https://github.com/unitaryai/detoxify/pull/47
• @gregpriday made their first contribution in https://github.com/unitaryai/detoxify/pull/54
• @anitavero made their first contribution in https://github.com/unitaryai/detoxify/pull/57
• @omidforoqi made their first contribution in https://github.com/unitaryai/detoxify/pull/68

Full Changelog: https://github.com/unitaryai/detoxify/compare/v0.5.0...v0.5.1

What's Changed

• freeze PL requirements by @Borda in https://github.com/unitaryai/detoxify/pull/40
• allow Load offline, expose local path by @Borda in https://github.com/unitaryai/detoxify/pull/39
• Pre-commit and import fixing by @Borda in https://github.com/unitaryai/detoxify/pull/36
• Updated pytorch-lightning version by @laurahanu in https://github.com/unitaryai/detoxify/pull/50

New Contributors

• @Borda made their first contribution in https://github.com/unitaryai/detoxify/pull/40

Full Changelog: https://github.com/unitaryai/detoxify/compare/v0.4.0...v0.5.0

• Updated the multilingual model weights used by Detoxify with a model trained on the translated data from the 2nd Jigsaw challenge (as well as the 1st). This model has also been trained to minimise bias and now returns the same categories as the unbiased model. New best AUC score on the test set: 92.11 (89.71 before).
• All detoxify models now return consistent class names (e.g. "identity_attack" replaces "identity_hate" in the original model to match the unbiased classes).

New improved weights for the multilingual Detoxify model trained on the translated data from the 2nd Jigsaw Challenge as well as from the 1st. Trained with same labels as the unbiased model.

• New improved unbiased model and updated data loaders to replicate
• tests to check torch.hub is loading the models
• script to convert saved weights to detoxify format

• New improved weights for the unbiased Detoxify model trained on the datasets provided by the 1st and 2nd challenges

Added lightweight checkpoints trained with Albert for original and unbiased models.