Ravikumar M. N.
Menu
Résumé
Selected work

Case study

Model adaptation with LLMs, embeddings, and classifiers

Fine-tuning where prompting and retrieval run out, using SFT with LoRA, contrastive embeddings, and transformer classifiers taken from 78% to 89% accuracy.

Role
Machine Learning Engineer
Context
Acranton · projects
When
2022 – 2025
Core stack
PyTorch, Hugging Face Transformers, PEFT / LoRA
sentiment accuracy, up from 78%
89%
inference latency
−30%
F1 on imbalanced classes
+10%

Problem

Off-the-shelf models were close but not good enough. They didn’t follow domain-specific instructions reliably, retrieval missed domain phrasing, and classifiers plateaued on the classes that mattered most. Each needed adaptation, under a GPU budget.

Constraints

  • Compute. Full fine-tuning of large models wasn’t affordable, so memory and training cost had to come down.
  • Production readiness. Models had to be served fast, not just score well offline.
  • Imbalanced labels. The minority classes were often the important ones.

Approach

LLMs: SFT with parameter-efficient adapters

I fine-tuned open-source LLMs for domain-specific instruction following using supervised fine-tuning (SFT) and PEFT. LoRA adapters kept the base model frozen, which reduced GPU memory and training cost while preserving model performance. The training pipelines covered dataset curation, prompt formatting, evaluation metrics, and inference optimization.

Embeddings: contrastive fine-tuning for retrieval

General-purpose embeddings blur domain distinctions, so I fine-tuned transformer embedding models on domain-specific text pairs with a contrastive objective (Sentence Transformers). This improved top-k retrieval relevance and downstream QA performance in RAG and semantic search.

Classifiers: accuracy and latency together

At Acranton I optimized and fine-tuned transformer NLP models, raising sentiment classification accuracy from 78% to 89% while cutting inference latency 30%. On imbalanced sentiment tasks, Hugging Face Transformers fine-tuning gave a 10% F1 improvement. Custom classification models raised data-pipeline accuracy 18% and cut processing time 25%.

Model adaptation pipelineCurated, formatted domain data feeds three kinds of adaptation: supervised fine-tuning of an LLM with LoRA adapters, contrastive fine-tuning of embedding models on domain text pairs, and fine-tuning of transformer classifiers. Each passes an evaluation gate on task metrics before optimized serving; evaluation results feed back into data curation.iterate on dataDomain datacurate · formatLLM · SFT + LoRAfrozen base, low-rank adaptersEmbeddings · contrastivedomain text pairsClassifier · fine-tune78% → 89% accuracyEvaluation gatetask metricsServingoptimized inference
Fig.One adaptation loop for three model types. Evaluation results feed back into data curation.

Results

sentiment classification accuracy, up from 78%
89%
inference latency
−30%
F1 on imbalanced sentiment classes
+10%
pipeline accuracy / processing time (custom classifiers)
+18% / −25%

Stack

Training
  • PyTorch
  • Hugging Face Transformers
  • PEFT
  • LoRA
  • Sentence Transformers
Experimentation
  • Weights & Biases
  • Optuna
  • MLflow