LLM Fine-Tuning & MLOps Pipeline

Standard pre-trained language models require domain adaptation and instruction-tuning to operate reliably on enterprise tasks. This project implements a robust, end-to-end LLM fine-tuning and MLOps evaluation pipeline designed to optimize open-source and proprietary backbones while ensuring performance reproducibility.

Parameter-Efficient Fine-Tuning (PEFT)

To optimize resource utilization, the pipeline implements Low-Rank Adaptation (LoRA) using Hugging Face Transformers. By freezing base weights and introducing trainable rank decomposition matrices, trainable parameter count was reduced by 95% without substantial performance degradation, making the training process highly accessible on consumer-grade hardware. The pipeline supports multi-stage instruction tuning and prompt template optimization.

MLOps, Versioning, and Tracking

Model metadata and hyperparameters are meticulously tracked using MLflow. Over 100+ runs were registered to evaluate learning rates, batch sizes, rank configurations, and prompt variations. Model weights are version-controlled using Data Version Control (DVC) backed by Git-based metadata tracking, ensuring full lineage auditing. Automated prompt testing evaluates LLM outputs across OpenAI, Anthropic, GLM, and local backbones.

Automated Evaluation with Ragas & TruLens

Rather than relying on ad-hoc inspections, output alignment is validated using structured frameworks: Ragas (for context precision, recall, and faithfulness) and TruLens (for prompt-response triads). This automated regression testing protects models from degradation during prompt updates or base model swap-outs.