v0.1.0 — Pretrained LLM Adaptation

Pretrained LLM Adaptation v0.1.0

End-to-end applied ML engineering project for adapting a pretrained language model to BANKING77 intent classification using DeBERTa-v3-small and LoRA/PEFT.

Included

• Pretrained model adaptation with LoRA
• Frozen-model and TF-IDF baselines
• LoRA rank ablation and model selection
• Separate learning rates for task head and adapter parameters
• Data validation, deduplication, and train/test leakage checks
• Accuracy, macro F1, weighted F1, classification report, and confusion matrix
• Error analysis and prediction comparison
• Reproducible training and evaluation configuration
• CLI inference support via llm-adapt
• Automated CI quality checks and ML smoke tests
• Automated benchmark and release-integrity workflow

Technical Stack

Python · PyTorch · Transformers · PEFT · Hugging Face Datasets · scikit-learn · Accelerate · pytest · Ruff · mypy

Scope

The project focuses on the engineering workflow required to adapt, evaluate, validate, and reproduce pretrained-model behavior rather than presenting a single training script or notebook.

See the repository README and docs/EXPERIMENT-METHODOLOGY.md for the complete methodology and reproducibility details.