Changelog¶
All notable changes to Mistral NER will be documented in this file.
The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.
[Unreleased]¶
Added¶
- Comprehensive MkDocs documentation with Material theme
- Support for 9 datasets including traditional NER and PII detection
- Multi-dataset training with three mixing strategies
- Advanced loss functions for handling class imbalance
- Hyperparameter optimization with Ray Tune and Optuna
- 8-bit and 4-bit quantization support
- WandB integration with offline mode support
- Validation script with comprehensive reporting
- GitHub Actions CI/CD pipeline
Changed¶
- Default quantization from 8-bit to 4-bit for better memory efficiency
- Enhanced LoRA configuration with more target modules
- Improved error handling and logging throughout
Fixed¶
- 8-bit quantization now works correctly when 4-bit is disabled
- Memory leak issues during long training runs
- Label alignment issues with subword tokenization
[0.2.0] - 2024-01-15¶
Added¶
- Initial public release
- Basic Mistral-7B fine-tuning for NER
- CoNLL-2003 dataset support
- LoRA and 8-bit quantization
- Basic training and inference scripts
Security¶
- No known security issues
For detailed changes, see the Git commit history.