Gradian

Gradian provides in-depth training-data attribution for LoRA fine-tuned language models. It helps pinpoint the exact training examples responsible for performance regressions in specific model capabilities. Key features include: • Capability regression evaluation • Training example attribution ranking • Dataset and configuration audits • Deterministic diagnostic checks This tool is designed to identify the data points that negatively impact model performance, helping engineers debug and refine their fine-tuning processes. It scores each training example against the lost capability and conducts a thorough examination of dataset and configuration issues that might be missed by purely mathematical influence calculations. Gradian offers a comprehensive view of what went wrong, from identifying truncated completions to flagging problematic hyperparameters. Gradian is ideal for engineering teams working with fine-tuned models who need to understand the root causes of performance drops. It enables more efficient model debugging, prevents catastrophic forgetting, and ensures that fine-tuning efforts lead to desired improvements without introducing new issues. The tool provides actionable insights, helping users resolve issues like overly aggressive learning rates or problematic dataset distributions.