Change Log¶
Last Updated: 2026-07-11 Version: v0.2.0
2025-10-16¶
- Added
src/modeling.py,src/logging_utils.py,src/data/datasets.py, andsrc/training/trainer.pyto implement the modular training stack. - Updated Hydra configs (
configs/base/default.yaml,configs/model/base.yaml,configs/training/base.yaml,configs/training/data/tiny.yaml) and added a sample TSV dataset atdata/tiny_text_classification.tsv. - Documented the trainer stack and reproducibility checklist in
README.md; refreshedsrc/data/__init__.pyexports and error logging notes.
2025-10-15¶
- Created
src/modeling.pyto centralise Hugging Face model/tokenizer loading with optional LoRA/PEFT hooks driven by Hydra config values. - Added
src/training/trainer.pyand exported classes to provide a mixed-precision aware trainer with evaluation, gradient accumulation, logging integration, and best-k checkpoint retention. - Introduced
src/logging_utils.pyto initialise TensorBoard/MLflow sessions and emit metrics from the trainer. - Added
src/data/datasets.pyfor TSV-based text classification datasets and DataLoader construction. - Restructured Hydra configs (
configs/base/default.yaml,configs/model/base.yaml,configs/training/base.yaml,configs/training/data/tiny.yaml) to compose defaults while preserving legacy aliases. - Documented the modular training stack and config layout updates in
README.mdand refreshed the model registry / reproducibility docs. - Added targeted unit tests for modeling, datasets, trainer, and logging utilities under
tests/to guard the new functionality.