Screen your fine-tuning data for a backdoor trigger phrase

You downloaded a fine-tuning dataset from a source you don't fully trust, and a handful of poisoned rows could plant a trigger nobody would ever notice.

Intermediate to advancedRead 7 minDoing: an afternoon, most of it in the frequency scan and the throwaway classifier8 steps

▶ Start

What you'll have at the end

A before/after comparison showing whether any tested trigger phrase reliably flips your fine-tuned model's behavior.

Before you start

You need

A downloaded fine-tuning dataset of short labeled text examples from a source you don't fully trust, and the ability to run a short local script that trains a small text classifier.

Not covered

Scanning a downloaded model checkpoint file for planted code, or checking an already-fine-tuned model for memorized training data; this only screens a labeled dataset before it goes anywhere near a real fine-tune.

Leans on

8 steps

Where to go next

Can your fine-tuned model be made to leak memorized training data?

go there instead once you've already fine-tuned on data you trust and want to check what the model might recite back, rather than screening a dataset before training even starts

Scan a checkpoint for a hidden pickle exploit

go there instead when what you downloaded is a model checkpoint file, not a training dataset; the risk lives in the file format, not in label frequencies

Checked 18 Aug 2026

Part of the LLM Security & Safety cookbook