Back to Projects
LLM Fine-tuning & Safety

Fine-tuned Dolphin LLM (Safety Research)

Performed memory-efficient fine-tuning on large models to test the effects of LLMs without guardrails in a university setting.

Research Objective

Understanding the boundaries of open-source LLMs through instruction-following optimization. This research focused on making models more direct and honest while analyzing the risks associated with removing standard safety alignment.

  • PEFT Optimization: Utilizing Parameter-Efficient Fine-Tuning (PEFT) to optimize models on limited hardware.
  • Data Alignment: Curating datasets to prioritize honesty and directness in model responses.
  • Impact Analysis: Testing model behavior in a controlled academic setting to measure accuracy vs. hallucination rates.