Efficient Model Adaptation — PEFT-VQA

This repository is currently focused on one project: a domain-specialized technical VQA assistant for computer vision, machine learning, and computer graphics.

Project Focus

Build a grounded assistant that answers technical questions by combining:

• retrieval-augmented generation (RAG)
• parameter-efficient fine-tuning (LoRA-first)
• quantized inference for efficient deployment

Core flow:

question -> retrieve evidence -> reason with adapted model -> grounded answer

Scope

• concept explanation and method comparison
• paper and loss-function summarization
• code/debug-oriented technical support
• practical recommendations (dataset, metric, model choices)

Primary Files

• Orientation + roadmap: AGENTS.md
• Narrative + backlog snapshot: docs/writeup.tex

Execution Principles

• measurable: fixed metrics and ablation matrix
• efficient: PEFT and quantization tradeoff tracking
• reproducible: versioned configs, checkpoints, and reports