LLM Debugger: Identifying Fine-Tuning Failures
Gradian is a training data attribution engine for LoRA fine-tuned LLMs, simplifying the process of identifying and correcting errors.
Gradian is a training data attribution engine designed for LoRA fine-tuned LLMs. It helps identify which training examples negatively impacted model capabilities during fine-tuning. By analyzing your training data, hyperparameters, and loss curves, it pinpoints the problematic examples through gradients. Gradian is intended to explain model behavior after fine-tuning, enabling users to enhance their training processes effectively.