Language model training without degeneration
A fine-tuning method masks selected tokens and uses likelihood-based learning to preserve the model’s language capabilities.
- First priority
- Applicant
- LG Management Development Institute

Patent records
- KR102934483B1
South Korea · Grant
- Filed
- Grant published
- US20260195547A1
United States · Published application
- Filed
- Published
- WO2025234866A1
International (PCT) · Published application
- Filed
- Published
- KR20260037080A
South Korea · Published application
- Filed
- Published
- KR20250161395A
South Korea · Published application
- Filed
- Published