Research · ICTA 2026

ThreadLearn — LLM Fine-tuning Research

Academic research: QLoRA fine-tuning of Qwen2.5-Coder-1.5B with a BM25 RAG pipeline of 2,055 documents to detect and fix Node.js race conditions. Reported results: 73.3% versus GPT-3.5-turbo at 65%, with a model 125 times smaller.

Technologies

• Python
• QLoRA
• RAG / BM25
• Qwen2.5-Coder
• HuggingFace
• FastAPI
• Docker