- Added Python 3.8 and 3.9 support and library requirements are optimized in 1525 and 1578.
- Removed all TensorFlow components and default trainer replaced with `torch_trainer` in 1574.
- Added Russian SuperGLUE models and submission generation in 1577.
- Added NER case-agnostig config in 1570.
- Added external metrics support in 1546.
- Nested config overwriting mechanism in 1561.
- Refactoring of the training logging in 1572.
- KBQA models migrated to PyTorch in 1569.
- Classification models migrated to PyTorch in 1565.
- NER models migrated to PyTorch in 1545.
- Context question answering models migrated to PyTorch in 1539.
- Entity Linking migrated to PyTorch and reduced RAM and VRAM consumption in 1516.
- Added config deprecation mechanism in 1547.
- `torch_bert_ranker` now uses the same Hugging Face Transformers version as the rest of the components in 1532.
- Models and components removed in 1523, 1526, 1534, 1533, 1537, 1544, 1560, 1563, 1564, 1573.
- Fixed a problem with pre-trained BERT models by DeepPavlov in 1502. Resolves 1275 and 1390.
- Fixed en_core_web_sm load error during tests in 1524.
- Removed Telegram, MSBot Framework, Yandex Alice and Amazon Alexa connectors in 1548.
- Documentation updated in 1517, 1529.