Gradian: Open-Source Tool to Pinpoint Poison Data Causing LLM Fine-Tuning Failures

vylara-ai · reddit · 2026-08-03

To solve the common pain point of models degrading silently during fine-tuning, a developer built Gradian, an open-source debugging tool.

Gradian computes per-example gradients over the LoRA adapter to rank and cluster training data, identifying exactly which examples erode specific capabilities. It also catches silent configuration errors like sequence truncation and missing EOS tokens. The author shared a counterintuitive finding: perfectly formatted poisoned data can be misjudged as helpful by standard gradients, requiring the subtraction of actual output gradients for accurate detection.

Original post →

More from coding & agent

coding & agent channel →